Artificial intelligence model monitoring method and wireless communication device

Abstract

Provided in the embodiments of the present application are an artificial intelligence model monitoring method and a wireless communication device. The artificial intelligence model monitoring method is executed by a first device, the first device is deployed with a channel state information (CSI) encoder that applies an artificial intelligence model, and a second device is correspondingly deployed with a CSI decoder that applies the artificial intelligence model. The method comprises: receiving second CSI sent by the second device; and comparing the second CSI with first CSI of the first device, wherein the second CSI is recovered CSI outputted by the CSI decoder of the second device, and the first CSI is original CSI inputted to the CSI encoder of the first device.

Description

Artificial intelligence model monitoring methods and wireless communication devices Technical Field

[0001] This application relates to the field of wireless communication technology, specifically to an artificial intelligence model monitoring method and a wireless communication device. Background Technology

[0002] In existing technologies, Artificial Intelligence / Machine Learning (AI / ML) is a system that can replace human labor through computational learning. AI / ML can be used to solve various problems, such as natural language processing, computation, and graphics processing. In recent years, AI / ML has been applied to the field of communications. However, there are unresolved issues regarding model monitoring methods for AI / ML applications in communications. Therefore, there is a need to propose an artificial intelligence model monitoring method and wireless communication devices to improve upon the problems and / or other issues of existing technologies.

[0003] Summary of the Invention

[0004] This application provides an artificial intelligence model monitoring method and a wireless communication device to improve the problems and / or other problems of the prior art.

[0005] This application provides an artificial intelligence model monitoring method, executed on a first device. The first device is equipped with a Channel State Information (CSI) encoder applying an artificial intelligence model, and a second device is correspondingly equipped with a CSI decoder applying an artificial intelligence model. The method includes: receiving second channel state information (CSI) sent by the second device; and comparing the second CSI with a first CSI from the first device. The second CSI is a recovered CSI output by the second device's CSI decoder, and the first CSI is the original CSI input by the first device's CSI encoder. This improves the reliability of the artificial intelligence model monitoring results, reduces the signaling overhead of the artificial intelligence model monitoring, and / or reduces the air interface load.

[0006] This application provides an artificial intelligence model monitoring method, executed on a first device. The first device is equipped with a Channel State Information (CSI) encoder or decoder that applies an artificial intelligence model, and a second device is correspondingly equipped with a CSI decoder or encoder that applies an artificial intelligence model. The method includes: receiving a second precoding matrix indication (PMI) sent by the second device; and comparing the second PMI with a first PMI of the first device; wherein the second PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the second device, and the first PMI is the corresponding PMI of the original CSI input by the CSI encoder of the first device; or the second PMI is the corresponding PMI of the original CSI input by the CSI encoder of the second device, and the first PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the first device.

[0007] By comparing the second PMI with the first PMI of the first device using the above technical solution, the reliability of the AI ​​model monitoring results can be improved, the signaling overhead of the AI ​​model monitoring can be reduced, and / or the air interface load can be reduced.

[0008] This application provides an artificial intelligence model monitoring method, executed on a first device. The first device is equipped with a Channel State Information (CSI) encoder or decoder that applies an artificial intelligence model, and a second device is correspondingly equipped with a CSI decoder or encoder that applies an artificial intelligence model. The method includes: obtaining an original CSI through CSI measurement; determining corresponding first downlink channel quality information based on the original CSI; receiving second downlink channel quality information sent by the second device, wherein the second downlink channel quality information is determined based on a recovered CSI; and comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device; or the original CSI is the original CSI input by the CSI encoder of the second device, and the recovered CSI is the recovered CSI output by the CSI decoder of the first device.

[0009] By employing the above technical solution, the model monitoring performance indicators are compared based on the first downlink channel quality information and the second downlink channel quality information. This improves the reliability of the AI ​​model monitoring results, reduces the signaling overhead of AI model monitoring, and / or lowers the air interface load.

[0010] This application provides an artificial intelligence model monitoring method, executed on a first device. The first device is equipped with a Channel State Information (CSI) encoder or decoder applying an artificial intelligence model, and a second device is correspondingly equipped with a CSI decoder or encoder applying an artificial intelligence model. The method includes: receiving second Channel State Information (CSI) sent by the second device; and comparing the second CSI with a first CSI from the first device. The second CSI is a recovered CSI output by the second device's CSI decoder, and the first CSI is the original CSI input by the first device's CSI encoder; or the second CSI is the original CSI input by the second device's CSI encoder, and the first CSI is the recovered CSI output by the first device's CSI decoder. By comparing the second CSI with the first CSI from the first device, the model monitoring performance indicators are compared. This can improve the reliability of the artificial intelligence model monitoring results, reduce the signaling overhead of the artificial intelligence model monitoring, and / or reduce the air interface load. This application provides an artificial intelligence model monitoring method, executed on a first device. The first device is equipped with a Channel State Information (CSI) encoder or decoder that applies an artificial intelligence model, and a second device is correspondingly equipped with a CSI decoder or encoder that applies an artificial intelligence model. The method includes: receiving a second precoding matrix indication (PMI) sent by the second device; and comparing the second PMI with a first PMI of the first device; wherein the second PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the second device, and the first PMI is the corresponding PMI of the original CSI input by the CSI encoder of the first device; or the second PMI is the corresponding PMI of the original CSI input by the CSI encoder of the second device, and the first PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the first device.

[0011] By comparing the second PMI with the first PMI of the first device using the above technical solution, the reliability of the AI ​​model monitoring results can be improved, the signaling overhead of the AI ​​model monitoring can be reduced, and / or the air interface load can be reduced.

[0012] This application provides an artificial intelligence model monitoring method, executed on a first device. The first device is equipped with a Channel State Information (CSI) encoder or decoder that applies an artificial intelligence model, and a second device is correspondingly equipped with a CSI decoder or encoder that applies an artificial intelligence model. The method includes: obtaining an original CSI through CSI measurement; determining corresponding first downlink channel quality information based on the original CSI; receiving second downlink channel quality information sent by the second device, wherein the second downlink channel quality information is determined based on a recovered CSI; and comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device; or the original CSI is the original CSI input by the CSI encoder of the second device, and the recovered CSI is the recovered CSI output by the CSI decoder of the first device.

[0013] By employing the above technical solution, the model monitoring performance indicators are compared based on the first downlink channel quality information and the second downlink channel quality information. This improves the reliability of the AI ​​model monitoring results, reduces the signaling overhead of AI model monitoring, and / or lowers the air interface load.

[0014] This application provides a wireless communication device, including a processor and a memory. The memory stores a computer program, and the processor calls and runs the computer program stored in the memory to perform the aforementioned wireless communication method.

[0015] The user equipment provided in this application includes a processor and a memory. The memory is used to store computer programs, and the processor is used to call and run the computer programs stored in the memory to perform the aforementioned wireless communication method.

[0016] The base station provided in this application includes a processor and a memory. The memory stores a computer program, and the processor calls and runs the computer program stored in the memory to perform the aforementioned wireless communication method.

[0017] The positioning management function entity provided in this application includes a processor and a memory. The memory is used to store computer programs, and the processor is used to call and run the computer programs stored in the memory to perform the aforementioned wireless communication method.

[0018] The network element provided in this application embodiment includes a processor and a memory. The memory is used to store computer programs, and the processor is used to call and run the computer programs stored in the memory to perform the aforementioned wireless communication method.

[0019] The chip provided in this application embodiment is used to implement the above-described wireless communication method.

[0020] Specifically, the chip includes a processor for retrieving and running a computer program from memory, causing a device equipped with the chip to perform the aforementioned wireless communication method.

[0021] The computer-readable storage medium provided in this application embodiment is used to store a computer program that causes a computer to perform the above-described wireless communication method.

[0022] The computer program product provided in this application includes computer program instructions that cause a computer to perform the above-described wireless communication method.

[0023] The computer program provided in this application embodiment, when run on a computer, causes the computer to execute the above-described method for wireless communication.

[0024] The above technical solutions can improve the reliability of AI model monitoring results, reduce the signaling overhead of AI model monitoring, and / or reduce the load on the air interface. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0026] Figure 1A is a schematic diagram of the measurement and feedback process of channel state information (CSI).

[0027] Figure 1B is a schematic diagram of a two-sided model framework and execution flow for CSI compression in Artificial Intelligence / Machine Learning (AI / ML).

[0028] Figure 1C is a schematic diagram of the monitoring process for a user equipment (UE) side model;

[0029] Figure 1D is a schematic diagram of the monitoring process of a network (NW) side model;

[0030] Figure 2 is a schematic diagram of a wireless communication system architecture provided in an embodiment of this application;

[0031] Figure 3 is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0032] Figure 4 is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0033] Figure 5A is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0034] Figure 5B is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0035] Figure 5C is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0036] Figure 6A is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0037] Figure 6B is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0038] Figure 7A is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0039] Figure 7B is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0040] Figure 8A is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0041] Figure 8B is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application;

[0042] Figure 9 is a schematic structural diagram of a wireless communication device provided in an embodiment of this application;

[0043] Figure 10 is a schematic structural diagram of a chip according to an embodiment of this application;

[0044] Figure 11 is a schematic block diagram of a wireless communication system provided in an embodiment of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0046] Within the framework of Massive Multiple-Input Multiple-Output (MIMO) communication systems, the measurement and feedback of Channel State Information (CSI) is one of the core technologies of the physical layer. Accurate measurement and feedback of CSI information can precisely reflect the channel environment of the physical downlink control channel (PDSCH) between the base station (gNB) and the user equipment (UE), helping the gNB to implement correct data modulation and coding methods. This is of great significance for improving the throughput and spectral efficiency of the PDSCH channel and enhancing the overall performance of the communication system.

[0047] Specifically, the overall process of CSI measurement and feedback involves the gNB sending a Channel State Information-Reference Signal (CSI-RS) to the UE. Based on the received CSI-RS, the UE measures the raw CSI information of the corresponding sub-band within the Bandwidth Part (BWP). This raw CSI information is typically the precoding matrix of each sub-band or the eigenvector resulting from the precoding matrix after Singular Value Decomposition (SVD). The UE then processes this raw CSI information using a specific algorithm, converting it into CSI feedback information. This feedback information is transmitted to the gNB via uplink information, such as Physical Uplink Shared Channel (PUSCH) or Physical Uplink Control Channel (PUCCH). The gNB recovers the raw CSI information from the received feedback information and modulates and encodes the downlink PDSCH channel data based on the recovered precoding matrix. This completes the CSI measurement and feedback process.

[0048] Figure 1A is a schematic diagram of the measurement and feedback process of channel state information (CSI). Referring to Figure 1A, the CSI measurement and feedback process includes at least one of the following steps: Step 1: The base station sends channel measurement configuration information to the UE; Step 2: The base station sends channel measurement reference information to the UE; Step 3: The UE feeds back CSI information to the base station; and Step 4: The base station sends downlink data with added precoding to the UE.

[0049] The measurement and feedback process of CSI is shown in Figure 1A. The raw CSI information consists of the precoding matrices of each sub-band in the BWP or the feature vectors after SVD decomposition. The data volume of the raw CSI information is enormous and cannot be transmitted through Layer 1 (L1) signaling (or Physical Layer). Furthermore, the amplitude and phase coefficients of the precoding matrices of each sub-band in the raw CSI information have high correlation in both the spatial and frequency domains, which forms the theoretical basis for compressing the raw CSI information. Therefore, the process by which the UE converts the raw CSI information into CSI feedback information involves compressing the raw CSI information into a smaller Precoding Matrix Indicator (PMI) for feedback reporting. This "specific method" corresponds to the various codebooks designed. From version 15 (Rel-15) to version 18 (Rel-18), the codebook design has continuously evolved, including the Rel-15 Type I codebook, the Rel-15 Type II codebook, the Rel-16 Enhanced Type II codebook, the Rel-17 Further Enhanced Type II codebook, the Rel-18 Enhanced Type II codebook for Coherent Joint Transmission (CJT), and the Rel-18 Enhanced Type II codebook for PMI prediction, among others. This continuous evolution of codebook design aims to more accurately recover the original CSI information and further reduce the feedback overhead of PMI, thereby achieving an optimal trade-off between system performance and signaling load.

[0050] Since the inception of the Rel-18 NR Air Interface Artificial Intelligence (AI) / Machine Learning (ML) research project, CSI feedback enhancement has been one of the key application scenarios studied within the AI / ML project. AI / ML-based CSI compression is one of the two sub-scenarios within CSI feedback enhancement. AI / ML-based CSI compression, through data acquisition and AI / ML model training, uncovers the correlation relationships between the original CSI information of each subband in the BWP across the spatial, frequency, and temporal domains. Based on the Type II codebook, it further compresses the overhead of CSI feedback and improves the recovery accuracy of CSI feedback. During the standard research and discussion process, various companies thoroughly explored and reached an agreement on the basic framework of AI / ML-based CSI compression, recognizing it as a two-sided model.

[0051] In the AI / ML-based CSI compression framework, a dual-side model means that it includes two AI / ML models, running on the UE and gNB sides respectively. The AI / ML model running on the UE side is called the encoder, also known as the CSI generation part of the AI / ML model. Its function is to compress the reported CSI information, generating compressed CSI data which is then reported to the gNB. The input to this model is the raw CSI information obtained after measurement (the precoding matrices of each subband in the BWP or the set of feature vectors after SVD decomposition), and the output is the compressed CSI data. Since the encoder's AI / ML model can be considered a black box, the compressed CSI data output by this model can be considered a string of bits with a certain amount of data, without a clear physical meaning. The AI / ML model running on the gNB side is called the decoder, also known as the CSI reconstruction part of the AI / ML model. Its role is to decompress the compressed CSI information reported back to the user and recover the original CSI information. Therefore, the encoder and decoder in CSI compression are inverse processes. These two AI / ML models can be trained jointly or separately.

[0052] Figure 1B is a schematic diagram of a two-sided model framework and execution flow for CSI compression in Artificial Intelligence / Machine Learning (AI / ML). Referring to Figure 1B, w represents the raw CSI information obtained by the UE after CSI measurement, which is the input of the encoder; This represents the recovered CSI information obtained after decompression by the gNB's decoder; it is the decoder's output. Therefore, w and The degree of similarity determines the system performance of AI / ML-based CSI compression. The higher the similarity, the closer the CSI information after compression and restoration of the AI / ML model is to the accurate CSI information measured by the UE, and the better the overall system performance of CSI compression. Conversely, the lower the similarity, the worse the overall system performance of CSI compression.

[0053] Model monitoring for AI / ML-based CSI compression is a crucial component of AI model lifecycle management (LCM). Once the encoder and decoder AI / ML models in AI / ML-based CSI compression have been trained and successfully deployed to their respective network elements, model monitoring needs to be performed periodically or event-triggered during the inference process. This allows the system to monitor the performance status of the AI / ML models at all times and determine the appropriate LCM operations based on the monitoring results, such as model activation, deactivation, switching, selection, and rollback.

[0054] For AI / ML-based CSI compression, model monitoring is also an important research topic discussed in RAN1. In the various standards meetings of RAN1, agreements have been reached on model monitoring for AI / ML-based CSI compression in the following two aspects.

[0055] (1) Intermediate Key Performance Indicators (KPIs): Intermediate KPIs refer to performance indicators that need to be calculated and statistically analyzed during model monitoring. These indicators directly reflect the system performance of the monitored AI / ML model. For CSI compression, there are two intermediate KPIs agreed upon in the standard: SGCS (Standardized Generalized Cosine Similarity) and NMSE (Normalized Mean Square Error). Their calculation formulas are as follows:

[0056] As can be seen from the above formula, the calculation of the two intermediate key performance indicators requires obtaining the original CSI information w and the CSI information recovered after CSI compression and decompression. However, referring to Figure 1B, it can be seen from the dual-sided model framework and execution flow of AI / ML-based CSI compression that w and Belonging to different network elements, the UE can obtain the original CSI information w through CSI measurement after receiving the CSI-RS reference signal, while The gNB obtains the data after CSI decompression by its decoder. Therefore, to calculate and statistically analyze the two intermediate key performance indicators, SGCS and NMSE, and to ensure the smooth execution of AI / ML model-based CSI compression model monitoring, it is necessary to design corresponding signaling and data transmission mechanisms that allow a network element (UE or gNB) to simultaneously obtain both w and

[0057] (2) Procedure of intermediate KPIs based monitoring: To enable AI / ML-based CSI compression model monitoring, corresponding signaling and data transmission methods need to be designed to ensure that intermediate KPIs can be calculated and statistically analyzed. Therefore, the monitoring process based on intermediate KPIs needs to be considered. The agreement reached in the RAN1 meeting currently includes three AI / ML model processes: Case 1: Network-side monitoring based on target CSI, and actual channel estimation related to CSI reports; Case 2: UE-side monitoring based on the output of the CSI reconstruction model; Case 3: UE-side monitoring based on the output of the UE-side CSI reconstruction model.

[0058] Figure 1C is a schematic diagram of the model monitoring process on the User Equipment (UE) side. Referring to Figure 1C, UE-side model monitoring refers to recovering the CSI information output by the decoder on the base station side, such as the gNB side. The downlink transmission is sent to the UE, where SGCS or NMSE is calculated and statistically analyzed, and the model monitoring results are reported to the gNB in ​​model monitoring mode. The flowchart of UE-side model monitoring is shown in Figure 1C. The operation steps of UE-side model monitoring include at least one of the following steps: Step 1: The gNB sends a CSI-RS reference signal to the UE. The UE measures the actual CSI information (i.e., raw CSI information) based on the reference signal, inputs it into the encoder for CSI compression, generates compressed CSI feedback information, and reports it to the gNB. Step 2: After receiving the reported CSI feedback message (compressed CSI information), the gNB inputs the CSI feedback information into the decoder for CSI decompression to recover the CSI information. Step 3: The gNB sends the recovered CSI information to the UE. The UE calculates the model monitoring index SGCS based on the received CSI information recovered by the gNB and the actual CSI information measured based on the reference signal. Step 4: The UE feeds back the model monitoring results to the gNB, which serves as the basis for the gNB to decide on AI / ML model LCM operations.

[0059] Figure 1D is a schematic diagram of a network (NW) side model monitoring process. Referring to Figure 1D, network-side model monitoring refers to the uplink transmission of the raw CSI information w measured by the UE based on the CSI-RS reference signal to the gNB, where the network side performs SGCS or NMSE calculation and statistics to obtain the model monitoring results. The flowchart of network-side model monitoring is shown in Figure 1D. The operation steps of network-side model monitoring include at least one of the following steps: Step 1: The gNB sends the CSI-RS reference signal to the UE, and the UE measures the actual CSI information (i.e., the raw CSI information) based on the reference signal. The UE inputs the CSI information into the encoder for CSI compression to generate compressed CSI feedback information. The compressed CSI feedback information and the actual CSI information are then reported to the gNB together. Step 2: After receiving the above information, the gNB inputs the compressed CSI feedback information into the decoder for decompression to recover the CSI information. The recovered CSI information is then used to calculate the model monitoring index SGCS with the reported actual CSI information. There are unresolved issues regarding model monitoring methods in AI / ML applications within the communications field. For example, AI / ML applications in communications model monitoring methods suffer from at least one of the following problems:

[0060] Problem: The uncompressed CSI information data is enormous and cannot be supported by L1 / L3 layer signaling.

[0061] Both the raw CSI information obtained from CSI measurements on the UE side and the CSI information recovered by the decoder on the gNB side contain the precoding matrices of each subband of the channel measurement or the set of feature vectors after SVD decomposition, resulting in a massive amount of data. Transmitting unquantized and uncompressed CSI information during model monitoring would significantly increase the air interface load on the downlink or uplink channels, potentially exceeding the capabilities of traditional L1 or L3 signaling. Therefore, the standard needs to consider the recovery of CSI information from downlink transmissions during AI / ML-based CSI compression model monitoring. Alternatively, the original CSI information w transmitted uplink can be quantized and compressed to reduce transmission resource overhead. Furthermore, interactive signaling needs to be designed to ensure that the UE and gNB have a consistent understanding of the quantized and compressed CSI information.

[0062] Problem: Quantization compression of uncompressed CSI information introduces quantization errors, affecting the reliability of model monitoring performance metrics.

[0063] In AI / ML-based CSI compression model monitoring, transmitting quantized and compressed CSI information introduces quantization errors, causing deviations between the calculated and statistical values ​​of SGCS and NMSE indicators and their actual values. This reduces the reliability of AI / ML-based CSI compression model monitoring. Therefore, after quantizing and compressing the CSI information to be transmitted during model monitoring, it is necessary to consider the deviations in the calculated model monitoring indicators caused by quantization errors to determine the reliability of the model monitoring results.

[0064] Problem: Monitoring of CSI compression model based on PDSCH channel signal-to-interference-plus-noise ratio (SINR) and Channel Quality Indicator (CQI).

[0065] Even after quantization and compression, CSI information still has a relatively large data volume, typically reaching several hundred or even thousands of bits, resulting in significant overhead for the air interface. It is advisable to consider using other parameters with low-overhead characteristics related to PMI to replace the channel precoding matrix in calculating statistical GSCS and NMSE (Normalized Mean Square Error). This would significantly reduce transmission overhead while ensuring the reliability of AI / ML-based CSI compression model monitoring.

[0066] Some embodiments of this application propose the following solutions to the problems mentioned above:

[0067] 1. This paper proposes a model monitoring scheme for directly compressing the raw CSI measured by the UE based on an AI / ML model. Two implementation schemes are proposed: non-codebook quantization compression and codebook quantization compression. The recovered CSI information output by the decoder on the network side (e.g., gNB side) is quantized, compressed, and then sent to the UE. The specific technical solution compresses the data from the following dimensions:

[0068] (1) Subband sampling: This involves using the frequency domain correlation between subbands to select a portion of the subbands to recover CSI information for model monitoring. For example, assuming there are 10 subbands in the BWP and they are consecutive in pairs, for each pair of adjacent subbands, only one subband's CSI information can be taken for model monitoring, reducing the total overhead by 50%.

[0069] (2) Polarized Antenna Sampling: For multi-polarized antennas, the spatial correlation between antennas (e.g., deterministic phase relationships) is utilized to select the recovered CSI information of antennas in a portion of the polarization directions for model monitoring. For example, assuming a dual-polarized antenna, only the recovered CSI information of antennas in one polarization direction is taken for model monitoring, which can reduce the total overhead by 50%.

[0070] (3) Quantization of CSI data: This includes amplitude information quantization and phase information quantization. For example, by using protocol rules, amplitude and phase data represented by floating-point numbers (32 bits) can be quantized to the 4-bit level, reducing the total overhead by 87.5%.

[0071] 2. Considering that the overhead of quantized and compressed CSI information is still relatively large, an indirect model monitoring scheme based on recovered CSI information is designed. This involves using indirect measurements such as SINR / CQI to judge the performance of the AI / ML-based CSI compression model. Specifically, the gNB uses the recovered CSI information output from the decoder to precode the reference signal (e.g., CSI-RS / demodulation reference signal, DMRS), and transmits the precoded reference signal downlink to the UE. The UE calculates the SINR / CQI of the PDSCH channel corresponding to the recovered CSI information output from the gNB-side decoder, and calculates its correlation with the SINR / CQI of the PDSCH channel obtained through CSI measurement. This correlation serves as an intermediate key performance indicator for monitoring the AI / ML-based CSI compression model.

[0072] 3. For the model monitoring scheme of CSI compression based on AI / ML on the network side, similarly, two implementation schemes are proposed: non-codebook quantization compression and codebook quantization compression. The original CSI information input from the encoder by the UE side is quantized and compressed and reported to the gNB. The quantization compression of the original CSI information adopts three similar methods, namely (1) sub-band sampling; (2) polarized antenna sampling; (3) quantization of CSI data. The main differences are in the uplink and downlink processes and signaling.

[0073] In summary, some embodiments of this application mainly study the design of AI / ML model monitoring schemes in CSI compression methods based on AI / ML models. This invention mainly studies the model monitoring problem of AI / ML two-sided models used for CSI compression. Since the measurement data and recovery data of the two-sided model are located on opposite sides of the network, model monitoring requires transmitting the measurement data, recovery data, or their statistical calculation values ​​(referred to as intermediate KPIs) over the network. The technical solutions of some embodiments of this application design a method for transmitting precoding matrix parameters and / or other related parameters, covering two parameter types: measurement and model recovery, achieving one or more of the following technical effects: (1) Recovering the precoding matrix as accurately as possible, improving the accuracy of statistical calculations of intermediate key performance indicators in AI / ML model monitoring, and ensuring the reliability of AI / ML model monitoring results in CSI compression. (2) Minimizing the overhead of necessary uplink or downlink signaling in AI / ML model monitoring, and reducing the load on the air interface. The solutions provided by some embodiments of this application support model monitoring based on AI / ML-compressed CSI, solving the problem of how to calculate model monitoring indicators for both raw CSI and recovered CSI on the same network element, thereby minimizing air interface load. Some embodiments of this application employ multiple effective methods to significantly reduce the overhead of recovering CSI transmission for network-side model monitoring and the overhead of raw CSI reporting for UE-side model monitoring, while ensuring the statistical reliability of model monitoring performance indicator calculations. Simultaneously, by designing corresponding signaling combinations, effective interaction of key parameters related to CSI compression between the network and UE is ensured, thereby guaranteeing consistency in the understanding of quantized compressed CSI data between the network-side and UE-side models. Finally, by designing an indicator parameter for measuring quantization error and a corresponding quantization reliability metric, it is ensured that the quantization error falls within the acceptable range of the model, thus avoiding model output distortion caused by over-quantization. The technical solutions of this application can be applied to various wireless communication systems, such as: Long Term Evolution (LTE) systems, LTE Frequency Division Duplex (FDD) systems, LTE Time Division Duplex (TDD) systems, 5G communication systems, or future wireless communication systems.

[0074] For example, the wireless communication system 100 used in this application embodiment is shown in FIG2. The wireless communication system 100 may include a base station 110, which may be a device communicating with user equipment (UE) 120. The base station 110 can provide communication coverage for a specific geographical area and can communicate with user equipment located within that coverage area. Optionally, the base station 110 may be an evolved Node B (eNB or eNodeB) in an LTE system, or it may be a mobile switching center, relay station, access point, vehicle-mounted equipment, wearable device, hub, switch, bridge, router, network-side equipment in a 5G network, or a base station in a future communication system, etc.

[0075] The wireless communication system 100 also includes at least one user equipment 120 located within the coverage area of ​​the base station 110. "User equipment" as used herein includes, but is not limited to, devices configured to receive / transmit communication signals via wired connections, such as via Public Switched Telephone Networks (PSTN), Digital Subscriber Line (DSL), digital cable, direct cable connection; and / or another data connection / network; and / or via a wireless interface, such as for cellular networks, Wireless Local Area Networks (WLAN), digital television networks such as DVB-H networks, satellite networks, AM-FM broadcast transmitters; and / or other user equipment. User equipment configured to communicate via a wireless interface may be referred to as a "wireless communication terminal," "wireless terminal," or "mobile terminal." Examples of mobile terminals include, but are not limited to, satellite or cellular phones; personal communications system (PCS) terminals that can combine cellular radiotelephone with data processing, fax, and data communication capabilities; PDAs that may include radiotelephones, pagers, Internet / intranet access, web browsers, notebooks, calendars, and / or Global Positioning System (GPS) receivers; and conventional laptop and / or handheld receivers or other electronic devices that include radiotelephone transceivers. User equipment can refer to access terminals, user units, user stations, mobile stations, mobile stations, remote stations, remote user equipment, mobile devices, wireless communication equipment, or user agents. Access terminals can be cellular phones, cordless phones, Session Initiation Protocol (SIP) phones, Wireless Local Loop (WLL) stations, personal digital assistants (PDAs), handheld devices with wireless communication capabilities, computing devices or other processing devices connected to a wireless modem, in-vehicle equipment, wearable devices, user equipment in 5G networks, or user equipment in future PLMN evolutions, etc.

[0076] The wireless communication system 100 also includes a network 130. Network 130 may be a Location Management Function (LMF) entity. Network 130 may be an IP mobile communication network operated by a mobile communication operator. For example, network 130 may be the core network used by a mobile communication operator to operate and manage the wireless communication system 100, or it may be the core network used by a virtual mobile communication operator such as an MVNO (Mobile Virtual Network Operator).

[0077] Network 130 can be connected to base station 110 as a relay device for transmitting user data. User equipment 120 sends and receives user data via network 130. It should be noted that user data communication is not limited to IP communication, but can also be non-IP communication.

[0078] In some embodiments of this application, the user equipment 120 is deployed with a channel state information (CSI) encoder using an artificial intelligence model, and the base station 110 is correspondingly deployed with a CSI decoder using an artificial intelligence model. The method includes: receiving second channel state information (CSI) sent by the base station 110; and comparing the second CSI with a first CSI of the user equipment 120; wherein the second CSI is a recovered CSI output by the CSI decoder of the base station 110, and the first CSI is the original CSI input by the CSI encoder of the user equipment 120.

[0079] In some embodiments of this application, the user equipment 120 is deployed with a channel state information (CSI) encoder using an artificial intelligence model, and the base station 110 is correspondingly deployed with a CSI decoder using an artificial intelligence model. The method includes: receiving a second precoding matrix indication (PMI) sent by the base station 110; and comparing the second PMI with a first PMI of the user equipment 120; wherein the second PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the base station 110, and the first PMI is the corresponding PMI of the original CSI input by the CSI encoder of the user equipment 120; or the second PMI is the corresponding PMI of the original CSI input by the CSI encoder of the base station 110, and the first PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the user equipment 120.

[0080] In some embodiments of this application, the user equipment 120 is deployed with a channel state information (CSI) encoder using an artificial intelligence model, and the base station 110 is correspondingly deployed with a CSI decoder using an artificial intelligence model. The method includes: obtaining an original CSI through CSI measurement; determining corresponding first downlink channel quality information based on the original CSI; receiving second downlink channel quality information sent by the base station 110, wherein the second downlink channel quality information is determined based on a recovered CSI; and comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information; wherein the recovered CSI is the recovered CSI output by the CSI decoder of the base station 110, and the original CSI is the original CSI input by the CSI encoder of the user equipment 120; or the original CSI is the original CSI input by the CSI encoder of the base station 110, and the recovered CSI is the recovered CSI output by the CSI decoder of the user equipment 120.

[0081] In some embodiments of this application, the user equipment 120 is deployed with a channel state information (CSI) encoder using an artificial intelligence model, and the base station 110 is correspondingly deployed with a CSI decoder using an artificial intelligence model. The method includes: receiving second channel state information (CSI) sent by the base station 110; and comparing the second CSI with a first CSI of the user equipment 120; wherein the second CSI is a recovered CSI output by the CSI decoder of the base station 110, and the first CSI is the original CSI input by the CSI encoder of the user equipment 120.

[0082] In some embodiments of this application, the user equipment 120 is deployed with a channel state information (CSI) encoder using an artificial intelligence model, and the base station 110 is correspondingly deployed with a CSI decoder using an artificial intelligence model. The method includes: receiving a second precoding matrix indication (PMI) sent by the base station 110; and comparing the second PMI with a first PMI of the user equipment 120; wherein the second PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the base station 110, and the first PMI is the corresponding PMI of the original CSI input by the CSI encoder of the user equipment 120; or the second PMI is the corresponding PMI of the original CSI input by the CSI encoder of the base station 110, and the first PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the user equipment 120.

[0083] In some embodiments of this application, the user equipment 120 is deployed with a channel state information (CSI) encoder using an artificial intelligence model, and the base station 110 is correspondingly deployed with a CSI decoder using an artificial intelligence model. The method includes: obtaining an original CSI through CSI measurement; determining corresponding first downlink channel quality information based on the original CSI; receiving second downlink channel quality information sent by the base station 110, wherein the second downlink channel quality information is determined based on a recovered CSI; and comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information; wherein the recovered CSI is the recovered CSI output by the CSI decoder of the base station 110, and the original CSI is the original CSI input by the CSI encoder of the user equipment 120; or the original CSI is the original CSI input by the CSI encoder of the base station 110, and the recovered CSI is the recovered CSI output by the CSI decoder of the user equipment 120.

[0084] This can improve the reliability of AI model monitoring results, reduce the signaling overhead of AI model monitoring, and / or reduce the load on the air interface.

[0085] Optionally, user equipment 120 can perform device-to-device (D2D) communication with each other.

[0086] Alternatively, 5G communication systems or 5G networks may also be referred to as New Radio (NR) systems or NR networks.

[0087] Figure 2 exemplarily illustrates a base station 110, two user equipments 120, and a network 130. Optionally, the wireless communication system 100 may include multiple base stations, and the coverage area of ​​each base station may include other numbers of user equipments; this embodiment does not limit this. Multiple base stations may be, for example, a first base station and a second base station. The first base station may be, for example, a source base station. The second base station may be, for example, a target base station. The network 130 may be a Location Management Function (LMF) entity. Optionally, the wireless communication system 100 may also include other network entities such as a network controller, a mobility management entity, and network elements; this embodiment does not limit this. For example, the network 130 may include other network entities such as a network controller, a mobility management entity, and network elements; this embodiment does not limit this.

[0088] It should be understood that devices with wireless communication functions in the network / system of this application embodiment can be referred to as wireless communication devices. Taking the wireless communication system 100 shown in FIG2 as an example, the wireless communication device may include a base station 110, a user equipment 120, and a network 130 with communication functions. The base station 110 and the user equipment 120 may be the specific devices described above, which will not be repeated here; the network 130 may be a Location Management Function (LMF) entity. The wireless communication device may also include other devices (network 130) in the wireless communication system 100. For example, the network 130 may include other network entities such as a network controller and a mobility management entity. This is not limited in this application embodiment.

[0089] It should be understood that the terms "system" and "network" are often used interchangeably in this document. The term "and / or" in this document merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Furthermore, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0090] To facilitate understanding of the technical solutions in the embodiments of this application, the technical solutions related to the embodiments of this application will be described below.

[0091] First embodiment: UE-side CSI compression model monitoring based on non-codebook transmission:

[0092] The first embodiment can be implemented independently. In some embodiments of this application, the first embodiment can also be combined with other embodiments, for example, the first embodiment can be implemented in combination with the second embodiment, the third embodiment, the fourth embodiment, the fifth embodiment, and / or the sixth embodiment. These embodiments can be implemented sequentially or in parallel, and this application is not limited thereto.

[0093] Figure 3 is a flowchart illustrating the artificial intelligence model monitoring method provided in this embodiment. As shown in Figure 3, the artificial intelligence model monitoring method is executed on a first device, which is equipped with a channel state information (CSI) encoder applying an artificial intelligence model. A second device is correspondingly equipped with a CSI decoder applying an artificial intelligence model. The method includes at least one of the following operations: Operation 301: Receiving second channel state information (CSI) sent by the second device. Operation 302: Comparing the second CSI with a first CSI from the first device. Wherein, the second CSI is the recovered CSI output by the CSI decoder of the second device, and the first CSI is the original CSI input by the CSI encoder of the first device.

[0094] In some embodiments of this application, the artificial intelligence model monitoring method is executed on a second device, the second device being equipped with a channel state information (CSI) decoder applying an artificial intelligence model, and a first device being equipped with a corresponding CSI encoder applying an artificial intelligence model. The method includes at least one of the following operations: sending a second channel state information (CSI) to the first device. The second CSI is a recovered CSI output by the CSI decoder of the second device. This allows the first device to compare the second CSI with its own first CSI.

[0095] By comparing the second CSI with the first CSI of the first device using the above technical solution, the reliability of the monitoring results of the artificial intelligence model can be improved, the signaling overhead of the artificial intelligence model monitoring can be reduced, and / or the air interface load can be reduced.

[0096] Specifically, the first device is, for example, the user equipment 120 or the base station 110 shown in Figure 2, and the second device is, for example, the corresponding base station 110 or user equipment 120 shown in Figure 2.

[0097] This first embodiment addresses the AI / ML-based CSI compression model monitoring mode on the UE side. It employs a non-codebook approach to quantize and compress the recovered CSI information output from the decoder. The quantized and compressed CSI information is then transmitted downlink to the UE via the PDSCH channel, where SGCS / NMSE calculations are performed on the UE side. This embodiment primarily addresses how to quantize and compress the recovered CSI information output from the decoder, and the design of downlink signaling to ensure the UE's consistent understanding of the quantized and compressed CSI data and its ability to correctly decode the recovered CSI information.

[0098] First, we need to evaluate the data size of the recovered CSI information output from the gNB-side decoder. Before quantization and compression, the formula for calculating the data size of the recovered CSI information is: S = 2N1N2vN3(Q A +Q P ).

[0099] Wherein, parameters N1 and N2 represent the antenna data in the horizontal and vertical directions on the base station side, while 2N1N2 represents the number of ports in the CSI-RS; parameter v represents the number of data layers, i.e., the Rank Indicator (RI) value in CSI measurement reporting; parameter N3 represents the number of sub-bands; parameter Q... A and Q PQ represents the bit length of the data type for the amplitude coefficient and phase coefficient of each element in the precoding matrix, respectively. For example, if the amplitude coefficient and phase coefficient of each element in the precoding matrix are both represented by single-precision floating-point numbers (FLOAT32), then Q... A =Q P = 32 bits. For example, assuming the number of CSI-RS ports is 32, the RI reported by CSI measurement is 2 (i.e., the number of data layers is 2), the number of subbands for CSI measurement in BWP is 10, and the data type of the amplitude coefficients and phase coefficients in the precoding matrix is ​​FLOAT32, then the amount of data of the recovered CSI information output by the decoder on the gNB side is 40 kbit.

[0100] In this embodiment, the quantization and compression of the recovered CSI information output by the decoder is performed in at least one of the following ways: Method 1: Select a portion of the recovered CSI from the sub-bands for transmission.

[0101] In some embodiments of this application, the recovered CSI is a recovered CSI after quantization compression based on a first quantization compression method, whereby the first quantization compression method involves selecting a portion of the recovered CSI and / or recovered differential CSI from sub-bands for transmission. In some embodiments of this application, selecting recovered differential CSI for transmission includes transmitting the recovered CSI of a certain sub-band and the recovered differential CSI of other sub-bands. In some embodiments of this application, the difference between the recovered CSI sub-band differential CSI of the other sub-bands and the recovered CSI of the certain sub-band is obtained. In some embodiments of this application, the recovered CSI sub-band differential information of all other sub-bands refers to: assuming there are N sub-bands with indexes 0, 1, ..., N-1, transmitting the recovered CSI information of sub-band n, and subtracting the recovered CSI information of sub-bands 1, ..., n-1, n+1, ..., N-1 from the recovered CSI information of sub-band n to obtain the recovered CSI sub-band differential information of sub-bands 1, ..., n-1, n+1, ..., N-1. In some embodiments of this application, selecting to transmit the recovery differential CSI includes: transmitting the difference values ​​sequentially, for example, transmitting 1, 2-1, 3-2, ..., N-(N-1) sequentially.

[0102] Specifically, since CSI information often exhibits high frequency domain correlation, meaning the precoding matrices of each sub-band have high similarity, it is unnecessary to calculate and statistically analyze w for each sub-band during model monitoring when calculating GSCS / NMSE monitoring performance metrics. The GSCS / NMSE allows for the selection of one or more representative subband precoding matrix coefficients for transmission and computation.

[0103] In some embodiments of this application, the first quantization compression method is to transmit the recovered CSI and / or the recovered differential CSI of the selected sub-bands based on rules or according to sub-band indicator instructions.

[0104] There are two specific implementation methods:

[0105] Alternative Solution 1: Rule-based.

[0106] In some embodiments of this application, selecting the recovery CSI and / or recovery differential CSI of the partial subbands for transmission based on the rules includes: selecting subbands whose channel quality conforms to preset rules. In some embodiments of this application, selecting the recovery CSI and / or recovery differential CSI of the partial subbands for transmission based on the rules includes: specifying that only the precoding matrix of the subband with the highest Channel Quality Indicator (CQI) is selected for calculation and statistics of model monitoring indicators. In some embodiments of this application, the preset rules are determined by one or more of the following methods: selecting all subbands exceeding a preset threshold; selecting N subbands exceeding the preset threshold; selecting K subbands with the highest channel quality, where K is greater than or equal to 1. The N subbands are subbands exceeding the preset threshold, and N can be greater than or equal to K. The N subbands can be numbered 0, 1, ..., N-1. The K subbands are the K subbands with the highest channel quality among the N subbands. In some embodiments of this application, when K equals 1, if multiple subbands have the same highest channel quality indicator value, only the subband corresponding to the first highest CQI in the reported CQI sequence sorted from left to right is selected, or only the subband corresponding to the first highest CQI in the high-order and low-order order of the reported CQI sequence is selected, or the subband is sorted by subband index, that is, only the subband with the largest or smallest subband index corresponding to the highest CQI in the reported CQI sequence is selected. In some embodiments of this application, the CQI sequence is included in the CSI report.

[0107] For example, it can be specified that only the precoding matrix of the subband with the highest CQI is selected for the calculation and statistics of the model monitoring index GSCS / NMSE. The rules for selecting subbands can be determined as follows: (1) a preset threshold is set, and all subbands exceeding the threshold are selected; (2) a preset threshold is set, and N subbands exceeding the threshold are selected; (3) the K subbands with the highest CQI are selected, where K is greater than or equal to 1. The N subbands are the subbands that exceed the preset threshold, and N can be greater than or equal to K. The indexes of the N subbands can be 0, 1, ..., N-1. The K subbands are the K subbands with the highest channel quality among the N subbands. Specifically, when using (3) and K=1, if multiple subbands have the same highest CQI value, only the subband corresponding to the first highest CQI in the left-to-right sorting of the reported CQI sequence is taken, or only the subband corresponding to the first highest CQI in the high-to-low byte sorting of the reported CQI sequence is taken, or sorted according to the subband index, that is, only the subband with the largest / smallest subband index corresponding to the highest CQI in the reported CQI sequence is taken. The CQI sequence is included in the CSI report reported by the UE. When using the traditional CSI report, it is included in the first part of the CSI report. When using the AI-oriented CSI report, if the AI-oriented CSI report uses the first part and the second part, it is included in the first part of the CSI report.

[0108] Alternative 2: Based on the sub-band indicator.

[0109] In some embodiments of this application, transmission based on the selection of the recovery CSI and / or recovery differential CSI of the selected sub-bands according to the sub-band indicator includes: selecting the precoding matrix of the sub-bands for downlink transmission and representing it with a sub-band indicator bitmap. This refers to the gNB automatically selecting the precoding matrix of the sub-bands for downlink transmission and representing it with a sub-band indicator bitmap. The sub-band indicator bitmap indicates the index of the selected sub-band. For example, if the BWP contains 10 sub-bands, and the gNB selects the precoding matrices of the 1st and 3rd sub-bands for model monitoring, then the corresponding sub-band indicator bitmap is represented as "1010000000". The gNB sends the sub-band indicator bitmap and the selected sub-band precoding matrix data to the UE. After receiving the information, the UE parses the sub-band indicator bitmap and calculates the model monitoring index SGCS / NMSE based on the selected sub-bands. The specific basis and method for the gNB to select sub-bands are implemented internally by the gNB; it can be random selection or selection based on certain rules or algorithms.

[0110] Method 2: Select the precoding coefficients corresponding to a single polarization direction for transmission.

[0111] In some embodiments of this application, the recovered CSI is a quantized and compressed CSI based on a second quantization compression method, whereby the precoding coefficients corresponding to a single polarization direction are selected for transmission. In some embodiments of this application, selecting precoding coefficients corresponding to a single polarization direction for transmission includes: the precoding matrix of the single polarization direction carries a polarization direction indicator, indicating the antenna polarization direction corresponding to the downlink transmission precoding matrix. In some embodiments of this application, the polarization direction indicator is represented by 1 bit; a polarization direction indicator of 0 indicates an antenna port in the +45° polarization direction, and a polarization direction indicator of 1 indicates an antenna port in the -45° polarization direction. In some embodiments of this application, the polarization direction indicator is represented by 1 bit; a polarization direction indicator of 1 indicates an antenna port in the +45° polarization direction, and a polarization direction indicator of 0 indicates an antenna port in the -45° polarization direction.

[0112] Specifically, in some embodiments of this application, the dimension of the precoding matrix in the CSI information is 2N1N2×v, that is, the number of rows in the matrix is ​​the number of CSI-RS ports, and the number of columns is the number of data layers. The upper N1N2 rows of the precoding matrix represent antenna ports in the +45° polarization direction, and the lower N1N2 rows represent antenna ports in the -45° polarization direction. In some embodiments of this application, the upper N1N2 rows of the precoding matrix represent antenna ports in the -45° polarization direction, and the lower N1N2 rows represent antenna ports in the +45° polarization direction. During model monitoring based on AI / ML CSI compression, the gNB can select one of the precoding matrix coefficients in one polarization direction for transmission. After receiving the above CSI information, the UE calculates w and in the corresponding polarization direction. The GSCS / NMSE is used. This halves the transmission overhead. Similar to Method 1, the precoding matrix for a single polarization direction in gNB downlink transmission needs to carry a polarization direction indicator, indicating the antenna polarization direction corresponding to the downlink transmission precoding matrix. In some embodiments of this application, the polarization direction indicator is represented by 1 bit, where 0 represents the antenna port in the +45° polarization direction (i.e., the upper N1N2 row of the complete precoding matrix), and 1 represents the antenna port in the -45° polarization direction (i.e., the lower N1N2 row of the complete precoding matrix). Method 2 is optional; in this case, the protocol may specify the use of a certain polarization direction.

[0113] Method 3: Quantize the amplitude and phase coefficients of the precoding matrix.

[0114] In some embodiments of this application, the recovered CSI is a recovered CSI after quantization compression based on a third quantization compression method, whereby the third quantization compression method quantizes the amplitude coefficients and / or phase coefficients of the precoding matrix. In some embodiments of this application, the amplitude coefficients and / or the phase coefficients are quantized using 3-bit or 4-bit quantization. In some embodiments of this application, quantizing the amplitude coefficients and / or the phase coefficients of the precoding matrix includes amplitude quantization using amplitude coefficient quantization indicators and / or amplitude quantization using phase coefficient quantization indicators. In some embodiments of this application, the amplitude coefficient quantization indicator is represented by 1 bit; an amplitude coefficient quantization indicator of 0 indicates that the amplitude coefficient is quantized using 3 bits, and an amplitude coefficient quantization indicator of 1 indicates that the amplitude coefficient is quantized using 4 bits. In some embodiments of this application, the phase coefficient quantization indicator is represented by 1 bit; a phase coefficient quantization indicator of 0 indicates that the phase coefficient is quantized using 3 bits, and a phase coefficient quantization indicator of 1 indicates that the phase coefficient is quantized using 4 bits.

[0115] In some embodiments of this application, the artificial intelligence model monitoring method further includes: receiving the quantization error of the recovered CSI before and after the third quantization compression method sent by the second device. Specifically, if the compression method of quantizing the amplitude coefficients and phase coefficients of the precoding matrix of the recovered CSI information in the third method is adopted, the resulting quantization error may affect the reliability of the calculation and statistics of the model monitoring performance index based on AI / ML CSI compression. Therefore, after quantizing the amplitude coefficients and phase coefficients of the precoding matrix, it is necessary to calculate the cosine similarity between the quantized precoding matrix and the precoding matrix recovered by the decoder output before quantization, and send the similarity to the UE.

[0116] Specifically, the coefficients of the elements of the unquantized precoding matrix are complex numbers, which can be represented as: The form is: where i, j represent the row and column numbers of the element coefficients in the precoding matrix, i = 1, 2, ..., 2N1N2, j = 1, ..., v, p ij θ represents the amplitude coefficient of the element. ij This represents the phase coefficient of the element. Generally, in CSI measurements, the obtained precoding matrix undergoes normalization, so the amplitude coefficient p... ij The range of values ​​for is 0 ≤ p ij ≤1, phase coefficient θ ij The range of values ​​for is 0 ≤ θ ijFor values ​​less than 2π, amplitude and phase coefficients are typically represented using floating-point numbers. Single-precision floating-point numbers require 32 bits, while double-precision floating-point numbers require 64 bits, resulting in significant overhead. Therefore, quantizing the amplitude and phase coefficients can greatly reduce data overhead.

[0117] In this embodiment, the amplitude coefficient can employ either 3-bit or 4-bit quantization, and an amplitude coefficient quantization indicator is set to indicate which quantization method is selected. This indicator is 1 bit, and can be set to 0 for 3-bit quantization and 1 for 4-bit quantization. The specific amplitude coefficient quantization method selected is determined by the gNB and then notified to the UE via the amplitude coefficient quantization indicator. The quantized value k of the amplitude coefficient and the amplitude coefficient p... ij The mapping relationships are shown in Table 1: (It is possible that the protocol only specifies one type.)

[0118] Table 1: k-element to p ij The mapping.

[0119] Similar to amplitude quantization, phase coefficient quantization typically employs either 3-bit or 4-bit quantization methods. The phase is quantized at equal intervals, and a phase coefficient quantization indicator is used to specify the selected quantization method. This indicator is 1 bit, and can be set to 0 for 3-bit quantization and 1 for 4-bit quantization. The specific phase coefficient quantization method selected is determined by the gNB and then communicated to the UE via the phase coefficient quantization indicator. If 3-bit phase quantization is used, then the quantized value c of the phase coefficient... ij With phase coefficient θ ij The mapping formula is:

[0120] If 4-bit phase quantization is used, then the quantized value c of the phase coefficient is... ij With phase coefficient θ ij The mapping formula is:

[0121] If the compression method of quantizing the amplitude and phase coefficients of the precoding matrix for recovering CSI information (Method 3) is adopted, quantization error will inevitably occur, which will affect the reliability of the calculation and statistics of model monitoring performance indicators based on AI / ML CSI compression. Therefore, after quantizing the amplitude and phase coefficients of the precoding matrix, it is necessary to calculate the cosine similarity between the quantized precoding matrix and the precoding matrix recovered by the decoder before quantization, and send this similarity to the UE as a basis for measuring the reliability of the GSCS / NMSE model monitoring performance indicators. If the cosine similarity is higher, it means that the quantization error generated after quantization is smaller, and the calculation and statistics of model monitoring performance indicators are more reliable.

[0122] Let's assume the precoded matrix recovered from the decoder's output is... The precoding matrix after amplitude and phase quantization and inverse transformation is: Then, the cosine similarity between them is calculated as follows:

[0123] Cosine similarity SGCS q Characterized the The magnitude of the quantization error after amplitude and phase quantization, SGCS q The larger the value, the smaller the quantization error. The greater the statistical reliability of the calculated performance metric SGCS / NMSE for model monitoring, the better. (SGCS is a cosine similarity metric.) q The quantized values ​​need to be sent to the UE so that the UE can measure the reliability of the model monitoring performance index calculation results and determine whether the performance index results of this model monitoring are usable.

[0124] Cosine similarity SGCS q The specific quantification method is as follows:

[0125] Cosine similarity SGCS q The value range is 0-1, but when SGCS q When the value is below a certain threshold, it indicates a large quantization error. The calculated performance metrics for model monitoring are unreliable, in which case SGCS... q The specific value of is meaningless and does not need to be quantized. Therefore, the specific quantization method is as follows: set a cosine similarity threshold S, and when SGCS q When the value is less than the threshold S, the quantization value is 0; when it is higher than the threshold S, proportional quantization is performed. This saves on cosine similarity SGCS. qThe number of bits. Proportional quantization: The original quantized value range is [a, b], the quantization step size is e, and the target quantized value is [x:1:y], then (ba) / 3 = y - x + 1. In this embodiment, the cosine similarity threshold S is set to 0.7, and the cosine similarity SGCS... q The portion higher than 0.7 is quantized using 4 bits, so the specific cosine similarity SGCS is... q The quantization mapping relationships are shown in Table 2:

[0126] Table 2: Cosine Similarity (SGCS) q The quantitative mapping relationship.

[0127] In some embodiments of this application, the recovered CSI includes first information (or a first part) and second information (or a second part), wherein the first information is a set of indicators for quantizing and compressing the recovered CSI, and the second information is CSI data for quantizing and compressing the recovered CSI. In some embodiments of this application, the first information is received through static indicators, periodic configuration, semi-static indicators, or dynamic indicators.

[0128] In some embodiments of this application, when the recovered CSI and / or the recovered differential CSI of the selected sub-band are transmitted based on rules, the data composition of the first information from the most significant bit (MSB) to the least significant bit (LSB) represents: quantization compression mode indicator: corresponding to whether the first quantization compression method, the second quantization compression method, and the third quantization compression method are used to quantize and compress the CSI information, wherein the first quantization compression method is to select the recovered CSI and / or the recovered differential CSI of the selected sub-band for transmission, the second quantization compression method is to select the precoding coefficients corresponding to a single polarization direction for transmission, and the third quantization compression method is to quantize the amplitude coefficients and / or phase coefficients of the precoding matrix; polarization direction selection indicator: indicating the antenna polarization direction corresponding to the downlink transmission precoding matrix; amplitude coefficient quantization indicator: indicating the quantization method of the amplitude coefficients of the precoding matrix elements; phase coefficient quantization indicator: indicating the quantization method of the phase coefficients of the precoding matrix elements; quantization compression similarity indicator: indicating the quantized value of the cosine similarity after quantization compression; second information data size indicator: indicating the number of data bits of the second information.

[0129] In some embodiments of this application, when the recovery CSI and / or recovery differential CSI of a portion of the subband is selected for transmission according to the subband indicator, the data composition of the first information from the most significant bit (MSB) to the least significant bit (LSB) represents: quantization compression mode indicator: corresponding to whether the first quantization compression method, the second quantization compression method, and the third quantization compression method are used to quantize and compress the CSI information, respectively. The first quantization compression method involves selecting the recovery CSI and / or recovery differential CSI of a portion of the subband for transmission, and the second quantization compression method involves selecting the precoding coefficients corresponding to a single polarization direction for transmission. The third quantization compression method quantizes the amplitude coefficients and / or phase coefficients of the precoding matrix; the polarization direction selection indicator indicates the antenna polarization direction corresponding to the downlink transmission precoding matrix; the amplitude coefficient quantization indicator indicates the quantization method of the amplitude coefficients of the precoding matrix elements; the phase coefficient quantization indicator indicates the quantization method of the phase coefficients of the precoding matrix elements; the quantization compression similarity indicator indicates the quantized value of the cosine similarity after quantization compression; the second information data size indicator indicates the number of data bits of the second information; and the subband indicator indicates the selected subband in a bitmap format.

[0130] In summary, UE-side CSI compression model monitoring based on non-codebook transmission requires the gNB to employ at least one of methods 1, 2, or 3 to quantize and compress the recovered CSI information output by the decoder, and then transmit the quantized and compressed CSI data downlink to the UE. To ensure that the UE can accurately parse the quantized and compressed CSI data, the entire downlink CSI data is transmitted using a first-part and second-part mode. The second part corresponds to the quantized and compressed CSI data recovered from the decoder output, while the first part is a set of various indicators for quantizing and compressing the recovered CSI information. The bit format of the first part is described below.

[0131] Case 1: The subband selection for CSI recovery uses a rule-based selection method, corresponding to Alternative Scheme 1. The first part of the data composition, from the most significant bit (MBS) to the least significant bit (LSB), represents:

[0132] Quantization compression mode indicator (3 bits): corresponds to whether to use mode 1, mode 2 and mode 3 to quantize and compress CSI information, where "0" means not to use and "1" means to use.

[0133] Polarization direction selection indicator (1 bit): Indicates the antenna polarization direction corresponding to the precoding matrix of the downlink transmission. If mode 2 is not used, this bit is meaningless.

[0134] Amplitude coefficient quantization indicator (1 bit): Indicates the quantization method of the amplitude coefficient of the precoding matrix element. If mode 3 is not used, this bit is meaningless.

[0135] Phase coefficient quantization indicator (1 bit): Indicates the quantization method of the phase coefficient of the precoded matrix element. If mode 3 is not used, this bit is meaningless.

[0136] Quantization Compression Similarity Indicator (4 bits): Indicates the cosine similarity (SGCS) after quantization compression. q The quantization value is meaningless if method 3 is not used.

[0137] The second part, the data size indicator (X-bit), indicates the number of data bits in the second part. The determination of X is as follows. As mentioned earlier, before quantization and compression, the formula for calculating the amount of data in the recovered CSI information is: S = 2N1N2vN3(Q A +Q P ), where parameters N1 and N2 represent the antenna data in the horizontal and vertical directions on the base station side, and 2N1N2 represents the number of CSI-RS ports; parameter v represents the data layer value; parameter N3 represents the number of sub-bands; parameter Q A and Q P These represent the bit lengths of the data types for the amplitude and phase coefficients of each element in the precoding matrix, respectively. For example, assuming the CSI-RS port count is 32, the RI reported by the CSI measurement is 2 (i.e., the data layer count is 2), the number of subbands used for CSI measurement in the BWP is 10, and the data type of the amplitude and phase coefficients in the precoding matrix is ​​FLOAT32, then the calculated uncompressed CSI information data size is 40960 bits. When implementing CSI compression, the goal is to reduce the data size by at least one order of magnitude, for example, to 4096 bits. In this case, we take 2N1N2 = 32, v = 2, N3 = 10, and Q... A =3, Q p =3, then the compressed CSI information data size is 3840, 3840 < 2 12 It can be indicated by X = 12 bits; if we take 2N1N2 = 64, v = 4, N3 = 10, Q A =4, Q p =4, then the compressed CSI information data size is 20480, 20480 < 2 15This can be indicated using X = 15 bits. Therefore, usually, setting X between 12 and 15 is sufficient to meet the requirements.

[0138] The size indication of the second part of the data size indicator can be indicated explicitly or implicitly. (1) Explicit indication: The size is directly indicated using X-bit; (2) Implicit indication: The size of the second part is calculated by the UE itself by indicating the number of ports, layers, sub-bands, and amplitude / phase quantization method. For example, in the example above, the number of ports is 64, the number of layers is 4, the number of sub-bands is 10, and the amplitude and phase quantization are 4 bits respectively. Theoretically, a total of 6+2+4+4+4=20 bits are needed. However, since the amplitude and phase quantization method has already been indicated in the amplitude coefficient quantization indicator and the phase coefficient quantization indicator, 8 bits can be saved, resulting in 6+2+4=12 bits, which is better than the explicit indication method. If the protocol stipulates that the number of layers and ports are not compressed (i.e., CSI information is not compressed using spatial correlation) and all are reported, then since the number of layers and ports of the transmitted signal is known to the UE, only the number of sub-bands needs to be indicated. In this example, only 4 bits are needed. Case 2: The subband selection for CSI recovery uses a selection method based on subband indicator, corresponding to Alternative Scheme 2. Therefore, the first part of the data composition needs to be supplemented by N3 bits compared to Case 1, representing the following from MSB to LSB:

[0139] Quantization compression mode indicator (3 bits): corresponds to whether to use mode 1, mode 2 and mode 3 to quantize and compress CSI information, where "0" means not to use and "1" means to use.

[0140] Subband Indicator (N3bit): The subband indicator bitmap indicates the index of the selected subband.

[0141] Polarization direction selection indicator (1 bit): Indicates the antenna polarization direction corresponding to the precoding matrix of the downlink transmission. If mode 2 is not used, this bit is meaningless.

[0142] Amplitude coefficient quantization indicator (1 bit): Indicates the quantization method of the amplitude coefficient of the precoding matrix element. If mode 3 is not used, this bit is meaningless.

[0143] Phase coefficient quantization indicator (1 bit): Indicates the quantization method of the phase coefficient of the precoded matrix element. If mode 3 is not used, this bit is meaningless.

[0144] Quantization Compression Similarity Indicator (4 bits): Indicates the cosine similarity (SGCS) after quantization compression. q The quantization value is meaningless if method 3 is not used.

[0145] The second part is the data size indicator (X-bit): Similar to Case 1, it can be indicated by (1) explicit indication or (2) implicit indication. Unlike Case 1, in the case of implicit indication, since Case 2 has already sent the N3-bit subband indicator, the UE only needs to count the non-zero elements in the subband indicator and does not need to indicate the number of subbands.

[0146] Subband Indicator (N3bit): Indicates the selected subband using a bitmap. As shown above, the number of bits in the first part of the data is fixed, and the UE can accurately parse the information in the first part using blind detection. Furthermore, the UE can determine the bit size of the second part based on the data size indicator of the second part in the first part, thereby accurately parsing the data in the second part. The first part of the data and the second part of the data can be transmitted separately or together. The first part of the data and the second part of the data can be transmitted via downlink control information (DCI) signaling, radio resource control (RRC) signaling, and / or medium access control-control element (MAC-CE) signaling.

[0147] The data in the first part above can be indicated in the following three ways:

[0148] Static indication: This refers to a protocol-defined or base station-pre-configured set of values ​​for the first part of the data, which applies to all the second part of the data. Periodic configuration: Configuring the first part of the data once can apply to multiple second parts of the data. The first part of the data can be transmitted via DCI, RRC messages, or system information block (SIB) messages. Simultaneously, the protocol specifies the length of the time window for the first part of the data to apply, or the associated second part of the data.

[0149] Semi-static indication: Configuring the first part of the data once can affect multiple second parts of the data. The first part of the data can be transmitted via RRC messages or SIB messages.

[0150] Dynamic indication: Configure the data in the first part to act on the associated data in the second part. The data in the first part can be transmitted via DCI or RRC messages. Alternatively, a set of data in the first part can be configured via RRC messages, and then a specific configuration can be activated via MAC-CE or DCI messages to act on the current data in the second part.

[0151] Second embodiment: UE-side CSI compression model monitoring based on codebook transmission:

[0152] The second embodiment can be implemented independently. In some embodiments of this application, the second embodiment can also be combined with other embodiments, for example, the second embodiment can be implemented in conjunction with the first embodiment, the third embodiment, the fourth embodiment, the fifth embodiment, and / or the sixth embodiment. These embodiments can be implemented sequentially or in parallel, and this application is not limited thereto.

[0153] Figure 4 is a flowchart illustrating the artificial intelligence model monitoring method provided in this embodiment. As shown in Figure 4, the artificial intelligence model monitoring method is executed on a first device, which is equipped with a Channel State Information (CSI) encoder applying an artificial intelligence model. A second device is correspondingly equipped with a CSI decoder applying an artificial intelligence model. The method includes at least one of the following operations: Operation 401: Receiving a second precoding matrix indication (PMI) sent by the second device. Operation 402: Comparing the second PMI with a first PMI of the first device. Wherein, the second PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the second device, and the first PMI is the corresponding PMI of the original CSI input by the CSI encoder of the first device.

[0154] In some embodiments of this application, the artificial intelligence model monitoring method is executed on a second device, the second device being deployed with a channel state information (CSI) decoder applying an artificial intelligence model, and a first device being correspondingly deployed with a CSI encoder applying an artificial intelligence model. The method includes at least one of the following operations: Operation 401: Sending a second precoding matrix indication (PMI) to the first device. Wherein, the second PMI is the corresponding PMI of the recovered CSI output by the second device's CSI decoder. This allows the first device to compare the second PMI with a first PMI of the first device.

[0155] By comparing the second PMI with the first PMI of the first device using the above technical solution, the reliability of the AI ​​model monitoring results can be improved, the signaling overhead of the AI ​​model monitoring can be reduced, and / or the air interface load can be reduced.

[0156] Specifically, the first device is, for example, the user equipment 120 or the base station 110 shown in Figure 2, and the second device is, for example, the corresponding base station 110 or user equipment 120 shown in Figure 2.

[0157] The second embodiment is for the AI / ML-based CSI compression model monitoring mode on the UE side. It adopts a codebook-based approach to compress the recovered CSI information output by the decoder into PMI. Then, the PMI after codebook compression is transmitted downlink to the UE through the PDSCH channel, and the SGCS / NMSE calculation and statistical operation is performed on the UE side.

[0158] In traditional codebook design schemes, many parameters highly related to PMI load are configured by higher-level parameters via RRC signaling. For example, the number L of spatial substrates selected in the Type II codebook of version 15 is configured by the higher-level parameter numberOfBeams (TS38.214, section 5.2.2.2.3); the enhanced Type II codebook of version 16 configures parameter combinations through the higher-level parameter paramCombination-r16. Selecting a given set of parameter combinations determines the number L of spatial substrates and the number M of frequency substrates. v And the maximum number of non-zero values ​​K0 in the spatial-frequency coefficient matrix (TS38.214 Table 5.2.2.2.5-1). Similarly, the Type II codebook further enhanced in version 17 also determines the corresponding parameters through the configuration parameter combination of the higher-level parameter paramCombination-r17 (TS38.214 Table 5.2.2.2.7-1); the Type II codebook enhanced in version 18 used for predicting PMI also determines the corresponding parameters through the configuration parameter combination of the higher-level parameter paramCombination-r18 (TS38.214 Table 5.2.2.2.10-1). After the above parameters are configured through the corresponding higher-level parameters, the UE cannot change the above parameters during CSI measurement and reporting. It can only measure and calculate the coefficients of each broadband component W1 and sub-band component W2 under the corresponding codebook under the constraints of the determined reporting codebook type, number of spatial bases, number of frequency bases, and other configuration parameters. However, for model monitoring applications based on AI / ML-based CSI compression, since this scenario uses a non-codebook CSI reporting mechanism (i.e., the CSI reported is a bit string compressed by the AI / ML encoder, not a PMI designed based on a codebook), there are no high-level parameters to configure as mentioned above. Therefore, this embodiment needs to process the recovered CSI information output by the decoder. When performing codebook-based quantization compression, the gNB needs to determine the codebook type corresponding to the generated PMI, the number of spatial basis selections L, and the number of frequency basis selections M. vThe parameters include the maximum number of non-zero coefficients K0 in the weighted amplitude coefficients of the sub-band components, and these parameters are sent to the UE so that the UE can correctly parse the PMI (carried in the second part, the base station codebook parameter indication has been known) and correctly recover the precoding matrix. In some embodiments of this application, the PMI corresponding to the recovered CSI includes first information (or may be called the first part) and second information (or may be called the second part). The first information includes configuration parameters related to codebook compression, and the second information is CSI data that has been quantized and compressed based on the codebook for the recovered CSI. In some embodiments of this application, the configuration parameters related to codebook compression include codebook type, the number of spatial basis selections, the number of frequency basis selections, and the maximum number of non-zero coefficients in the weighted amplitude coefficients of the sub-band components. In some embodiments of this application, the PMI corresponding to the recovered CSI corresponds to the parameter combination of the codebook type. In some embodiments of this application, the parameters of the PMI corresponding to the recovered CSI are indicated in the first information. In some embodiments of this application, the first information is received through static indication, periodic configuration, semi-static indication, or dynamic indication.

[0159] Specifically, the data format design for downlink CSI data is as follows:

[0160] The entire downlink CSI data is transmitted using a first-part and second-part transmission pattern. The second part corresponds to the PMI data, which is the recovered CSI information from the decoder output after codebook-based quantization and compression. The first part includes configuration parameters related to codebook compression, including codebook type, the number of spatial basis selections (L), and the number of frequency basis selections (M). v The maximum number of non-zero coefficients, K0, in the weighted amplitude coefficients of the sub-band components. Of course, the above parameters can also be determined by parameter combinations. For version 16eType II codebooks, version 17FeType II codebooks, and version 18eType II codebooks used for PMI prediction, tables of parameter combinations are defined. The above parameters are determined by specifying the index (row number) of the corresponding parameter combination table. In this embodiment, the PMI data after codebook-based quantization and compression of the recovered CSI information output by the decoder can follow the parameter combinations of the corresponding codebook type, or it can not follow the parameter combinations. The above parameters are determined by the gNB itself and indicated to the UE in the first part. The bit format of the first part is described below for these two cases respectively.

[0161] Case 1: Follow the parameter combination of the corresponding codebook type.

[0162] This case 1 only applies to codebooks of version 16e Type II and version 17Fe Type II. The first part consists of 21 bits of data, with the MSB to LSB representing (the protocol may only specify one type):

[0163] Codebook type indicator (2 bits): This indicates the codebook type of the PMI data after the recovered CSI information output by the decoder is quantized and compressed based on the codebook. The specific quantization values ​​are shown in Table 3.

[0164] Table 3: Codebook type indication.

[0165] Parameter Combination Index Indicator (3 bits): Used to indicate the index of the parameter combination for the corresponding codebook type. If the codebook type is version 16eType II, the corresponding parameter combinations are shown in Table 5.2.2.2.5-1 of TS38.214; if the codebook type is version 17FeType II, the corresponding parameter combinations are shown in Table 5.2.2.2.7-1 of TS38.214. Both tables only specify 8 parameter combinations, so the parameter combination index indicator can be represented by 3 bits.

[0166] Quantization compression similarity indicator (4 bits): Indicates the quantized value of the cosine similarity after quantization compression.

[0167] The second part's data size indicator (X-bit): The second part's data size indicator indicates the number of data bits in the second part. (1) Explicit indication: Directly indicates the number of PMI data bits in the second part. (2) Implicit indication: Since the parameter combination of the codebook is already indicated in the parameter combination index indicator (3 bits), the UE can obtain it itself according to the protocol. Therefore, it is only necessary to indicate the number of sub-bands selected so that the UE can calculate the size of the second part itself.

[0168] Case 2: Parameter combination that does not conform to codebook type.

[0169] This Case 2 applies to codebook types I, version 15 type II, version 16e type II, and version Fe type II. The first part consists of 18 bits of data, with the MSB to LSB representing:

[0170] Codebook type indicator (2 bits): Used to indicate the codebook type of the PMI data after the recovered CSI information from the decoder output is quantized and compressed based on the codebook.

[0171] Spatial basis number indication (2 bits): In the Type II codebooks of versions 15, 16e, and 17Fe, the configuration of the number L of spatial basis selections is involved. For the Type II codebook of version 15, L ∈ {2, 3, 4}, and for the Type II and Fe codebooks of version 16, L ∈ {2, 4, 6}. Therefore, the number of spatial basis selections can be 2, 3, 4, or 6, which can be indicated using 2 bits. Specific quantization values ​​are shown in Table 4.

[0172] Table 4: Indication of the number of airspace bases.

[0173] Frequency domain substrate selection percentage indicator (2 bits): M of the number of frequency domain substrates involved in the version 16e Type II codebook v The configuration, and the number of frequency domain substrates selected is proportional to the number of sub-bands N3, the conversion formula is as follows: Where parameter p v This represents the proportion of the number of frequency domain basis selections. In Table 5.2.2.2.5-1 of parameter combination TS38.214, There are three possible values, which can be indicated using 2 bits. See Table 5 for specific quantization values:

[0174] Table 5: Percentage of Frequency Domain Substrates Selected

[0175] Quantization compression similarity indicator (4 bits): Indicates the quantized value of the cosine similarity after quantization compression.

[0176] The second part's data size indicator (X-bit): The number of PMI data bits in the second part. (1) Explicit indicator: Directly indicates the number of PMI data bits in the second part. (2) Implicit indicator: Since the codebook type, number of spatial bases, and number of frequency bases are indicated in the previous indicators, it is only necessary to indicate the number of selected sub-bands and the maximum number K0 of non-zero coefficients in the weighting amplitude coefficients of the corresponding sub-band components, so that the UE can calculate the size of the second part itself. The number of bits in the first part of the data is fixed. The UE can accurately parse the information in the first part through blind detection and further accurately parse the data in the second part. The data in the first part and the data in the second part can be transmitted separately or together. The data in the first part and the data in the second part can be transmitted through DCI signaling, RRC signaling, and MAC-CE signaling.

[0177] The data in the first part above can be indicated in the following three ways:

[0178] Static indication: This refers to a set of values ​​for the first part of the data specified in the protocol or pre-configured by the base station, which applies to all data in the first part. Periodic configuration: Configuring the first part of the data once can apply to multiple second parts of the data. The first part of the data can be transmitted via DCI, RRC, or SIB messages. Simultaneously, the protocol specifies the length of the time window for the first part of the data to apply, or the length of the second part of the data associated with the first part of the data.

[0179] Semi-static indication: Configuring the first part of the data once can affect multiple parts of the data. The first part of the data can be transmitted via RRC messages or SIB messages.

[0180] Dynamic indication: Configure the data in the first part to affect the associated data in the second part. The first part of the data can be transmitted via DCI or RRC messages. Alternatively, a set of first part data can be configured via RRC messages, and then a specific configuration can be activated via MAC-CE or DCI messages to affect the current second part of the data.

[0181] Third embodiment: UE-side CSI compression model monitoring based on PDSCH channel SINR measurement:

[0182] The third embodiment can be implemented independently. In some embodiments of this application, the third embodiment can also be combined with other embodiments, for example, the third embodiment can be implemented in combination with the first embodiment, the second embodiment, the fourth embodiment, the fifth embodiment, and / or the sixth embodiment. These embodiments can be implemented sequentially or in parallel, and this application is not limited thereto.

[0183] Figure 5A is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application. As shown in Figure 5A, the artificial intelligence model monitoring method is executed on a first device, which is equipped with a Channel State Information (CSI) encoder applying an artificial intelligence model, and a second device is equipped with a CSI decoder applying an artificial intelligence model. The method includes at least one of the following operations: Operation 501A: Obtaining the original CSI through CSI measurement, and determining the corresponding first downlink channel quality information based on the original CSI. Operation 502A: Receiving the second downlink channel quality information sent by the second device. The second downlink channel quality information is determined based on the recovered CSI. Operation 503A: Comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device.

[0184] In some embodiments of this application, the artificial intelligence model monitoring method is executed on a second device. The second device is equipped with a Channel State Information (CSI) decoder that applies an artificial intelligence model, and the first device is equipped with a CSI encoder that applies an artificial intelligence model. The method includes at least one of the following operations: sending second downlink channel quality information to the first device. The second downlink channel quality information is determined based on the recovered CSI. This allows the first device to compare model monitoring performance metrics based on the first downlink channel quality information and the second downlink channel quality information. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device.

[0185] In some embodiments of this application, the artificial intelligence model monitoring method is executed on a second device, the second device being equipped with a channel state information (CSI) decoder applying an artificial intelligence model, and the first device being equipped with a corresponding CSI encoder applying an artificial intelligence model. The method includes at least one of the following operations: receiving first downlink channel quality information sent by the first device, wherein the first downlink channel quality information is determined based on the original CSI; comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information of the second device, wherein the second downlink channel quality information is determined based on the recovered CSI. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device.

[0186] By employing the above technical solution, the model monitoring performance indicators are compared based on the first downlink channel quality information and the second downlink channel quality information. This improves the reliability of the AI ​​model monitoring results, reduces the signaling overhead of AI model monitoring, and / or lowers the air interface load.

[0187] Specifically, the first device is, for example, the user equipment 120 or the base station 110 shown in Figure 2, and the second device is, for example, the corresponding base station 110 or user equipment 120 shown in Figure 2.

[0188] In some embodiments of this application, the first downlink channel quality information and the second downlink channel quality information include signal-to-interference-plus-noise ratio (SINR), reference signal received power (RSRP), reference signal received quality (RSRQ), received signal strength indicator (RSSI), or channel quality indicator (CQI). In some embodiments of this application, the difference between the first downlink channel quality information and the second downlink channel quality information is greater than a preset threshold. In some embodiments of this application, comparing the model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information includes comparing the similarity between the first downlink channel quality information and the second downlink channel quality information. In some embodiments of this application, comparing the model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information includes comparing the cosine similarity between the first downlink channel quality information and the cosine similarity between the original CSI and the recovered CSI. In some embodiments of this application, the cosine similarity and / or normalized mean square error between the first downlink channel quality information and the second downlink channel quality information are used as the model monitoring performance indicators.

[0189] This third embodiment proposes a method for UE-side CSI compression model monitoring based on SINR measurement of the PDSCH channel. In the execution of AI / ML-based CSI compression model monitoring, the calculation and statistics of GSCS / NMSE model monitoring performance indicators are completed on the UE side. However, unlike the previous two embodiments, the model monitoring performance indicators to be calculated and statistically analyzed in this embodiment are the SINR of each sub-band in the PDSCH channel calculated by the UE side based on the raw CSI information obtained by CSI measurement from the CSI-RS reference signal, and the SINR of each sub-band in the PDSCH channel calculated by the recovered CSI information output by the decoder on the gNB side (denoted as SINR). It is based on comparisons between ( ). Specifically, there are the following two implementation methods.

[0190] Alternative Solution 1: Differential Thresholding Method

[0191] In the two SINRs mentioned above, the signal-to-interference-plus-noise ratio (SINR) of the sub-band calculated from the original CSI information w obtained by CSI measurement from the input of the encoder on the UE side can be obtained by the CSI measurement module on the UE side.

[0192] gNB in ​​obtaining Then, the reference signal is... Precoding is performed, and the precoded reference signal is sent to the UE. The UE receives the reference signal. After precoding the reference signal, the following can be calculated: Corresponding signal-to-interference-plus-noise ratio Specific reference signals include CSI-RS and DMRS, both of which can achieve the above functions.

[0193] Since the reference signal, after proper precoding, theoretically eliminates the adverse effects of the channel on the signal, appropriate precoding will be reflected in the improvement of the SINR of the received signal. Therefore, the expected... Therefore, when performing model monitoring, we can either pre-configure the protocol or configure a threshold T in real time at the base station. or If the CSI performance is good enough, then the CSI compression model performance is considered to meet the requirements.

[0194] Alternative Solution 2: Precoding Performance Comparison Method

[0195] Because the SINR of the precoded subband is affected by the PM (precoding matrix), it can be determined by comparing the PM-precoded subbands corresponding to w. and by Corresponding Precoded subband The similarity between w and is used to measure the accuracy of CSI information recovery. Therefore, w and are used to measure the accuracy of CSI information recovery. The calculated SINR of each subband of the PDSCH channel and The cosine similarity between them is positively correlated with w and The cosine similarity between them can be used to compare SINR and The cosine similarity and normalized mean square error between the models are used as performance metrics for monitoring AI / ML-based CSI compression models. Their calculation formulas are as follows:

[0196] In the two SINR scenarios mentioned above, the UE calculates the precoding matrix PM according to the protocol specifications based on the raw CSI information w obtained through CSI measurement. Using this matrix PM, the UE performs post-coding on the un-precoded received CSI-RS signal (to simulate the network-side precoding operation and avoid the overhead of retransmitting measurements after the network precodes the CSI-RS with PM). Then, the post-coded received CSI-RS signal is used for CSI estimation and subband calculation. The calculation is the same as in Alternative Solution 1.

[0197] Case 1: Referring to Figure 5B, the CSI-RS reference signal is... Precoding. SINR can also be replaced by similar measurement calculations, such as RSRP, RSRQ, RSSI, etc. of the reference signal. Base stations, such as gNBs, use CSI-RS of the reference signal... Precoding is performed, and the precoded reference signal CSI-RS is sent to the UE. The UE receives the signal and... After precoding the reference signal CSI-RS, the following can be calculated based on the reference signal CSI-RS: Corresponding signal-to-interference-plus-noise ratio

[0198] Case 2: Referring to Figure 5C, the DMRS reference signal is... Precoding. SINR can also be replaced by similar measurement calculations, such as RSRP, RSRQ, RSSI, etc. of the reference signal DMRS. Base stations, such as gNBs, use the reference signal DMRS... Precoding is performed, and the precoded reference signal DMRS is sent to the UE. The UE receives the signal and... After precoding the reference signal DMRS, the following can be calculated based on the reference signal DMRS: Corresponding signal-to-interference-plus-noise ratio

[0199] Fourth Implementation Example: UE-side CSI compression model monitoring based on PDSCH channel CQI:

[0200] The fourth embodiment can be implemented independently. In some embodiments of this application, the fourth embodiment can also be combined with other embodiments, for example, the fourth embodiment can be implemented in combination with the first embodiment, the second embodiment, the third embodiment, the fifth embodiment, and / or the sixth embodiment. These embodiments can be implemented sequentially or in parallel, and this application is not limited thereto.

[0201] The fourth embodiment proposes a method for UE-side CSI compressed model monitoring based on CQI measurement of the PDSCH channel. In the execution of AI / ML-based CSI compressed model monitoring, the calculation and statistics of GSCS / NMSE model monitoring performance indicators are completed on the UE side. However, unlike the previous two embodiments, the model monitoring performance indicators to be calculated and statistically analyzed in this fourth embodiment are the CQI of each sub-band in the PDSCH channel calculated by the UE side based on the raw CSI information obtained by CSI measurement from the CSI-RS reference signal, and the CQI of each sub-band in the PDSCH channel calculated by the CSI information recovered from the decoder output by the gNB side (denoted as CQI). It is based on comparisons between ( ). Specifically, there are the following two implementation methods.

[0202] Alternative Solution 1: Differential Thresholding Method

[0203] In the two CQIs mentioned above, the signal-to-interference-plus-noise ratio (SIR) of the sub-band calculated from the raw CSI information w obtained by CSI measurement on the UE side encoder input can be obtained by the CSI measurement module on the UE side.

[0204] gNB in ​​obtaining Then, the reference signal is... Precoding is performed, and the precoded reference signal is sent to the UE. The UE receives the reference signal. After precoding the reference signal, the following can be calculated: Corresponding signal-to-interference-plus-noise ratio Specific reference signals include CSI-RS and DMRS, both of which can achieve the above functions.

[0205] Since the reference signal, after proper precoding, theoretically eliminates the adverse effects of the channel on the signal, appropriate precoding will be reflected in the channel quality (characterized by CQI) of the received signal. Therefore, the expected... Therefore, when performing model monitoring, we can either pre-configure the protocol or configure a threshold B in real time at the base station. If the performance of the precoding matrix obtained using the original CSI information and the recovered CSI information is similar, then the performance of the CSI compression model meets the requirements.

[0206] Alternative Solution 2: Precoding Performance Comparison Method

[0207] Because the channel quality of the precoded subband (characterized by CQI) is affected by the PM (precoding matrix), it can be determined by comparing the PM-precoded subbands corresponding to w. and by Corresponding Precoded subband The similarity between w and [other values] is used to measure the accuracy of CSI information recovery. Therefore, w and [other values] [are used to measure the accuracy of CSI information recovery]. The calculated subbands of the PDSCH channel and The cosine similarity between them is positively correlated with w and The cosine similarity between them can be used to... and The cosine similarity and normalized mean square error between the models are used as performance metrics for monitoring AI / ML-based CSI compression models. Their calculation formulas are as follows:

[0208] Or, more simply, using the absolute difference comparison method, if If ε is a relatively small value specified in the protocol, it is also considered to meet the model performance requirements.

[0209] In the two SINR scenarios mentioned above, the UE calculates the precoding matrix PM according to the protocol specifications based on the raw CSI information w obtained through CSI measurement. Using this matrix PM, the UE performs post-coding on the un-precoded received CSI-RS signal (to simulate the network-side precoding operation and avoid the overhead of the NW retransmitting the measurement after precoding the CSI-RS with PM). Then, the post-coded received CSI-RS signal is used for CSI estimation and subband calculation. The calculation is the same as in Alternative Solution 1.

[0210] Case 1: Referring to Figure 6A, the CSI-RS reference signal is... Precoding. gNB in ​​acquiring Then, the reference signal CSI-RS was used with Precoding is performed, and the precoded reference signal CSI-RS is sent to the UE. The UE receives the signal and... After precoding the reference signal CSI-RS, the following can be calculated based on the reference signal CSI-RS: Corresponding signal-to-interference-plus-noise ratio

[0211] Case 2: Referring to Figure 6B, the DMRS reference signal is... Precoding. gNB in ​​acquiring Then, the reference signal DMRS is... Precoding is performed, and the precoded reference signal DMRS is sent to the UE. The UE receives the signal and... After precoding the reference signal DMRS, the following can be calculated based on the reference signal DMRS: Corresponding signal-to-interference-plus-noise ratio

[0212] Note: CQI calculation involves four tables in the standard (TS38.214 Tables 5.2.2.2.1-2, 5.2.2.2.1-3, 5.2.2.2.1-4, and 5.2.2.2.1-5). The specific table selected by the UE during CQI calculation is configured by the higher-layer parameter CSI-ReportConfig::cqi-Table. In AI / ML-based CSI compression model monitoring applications, to ensure the UE's understanding of CQI and... Consistent understanding, in calculation When doing so, you need to ensure that you use the same table as when calculating CQI.

[0213] Fifth Implementation Example: Monitoring of Network-Side CSI Compression Model Based on Non-Codebook Transmission:

[0214] The fifth embodiment can be implemented independently. In some embodiments of this application, the fifth embodiment can also be combined with other embodiments, for example, the fifth embodiment can be implemented in combination with the first embodiment, the second embodiment, the third embodiment, the fourth embodiment, and / or the sixth embodiment. These embodiments can be implemented sequentially or in parallel, and this application is not limited thereto.

[0215] Figure 7A is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application. As shown in Figure 7A, the artificial intelligence model monitoring method is executed on a second device. The second device is equipped with a channel state information (CSI) decoder that applies an artificial intelligence model, and the first device is equipped with a CSI encoder that applies an artificial intelligence model. The method includes at least one of the following operations: Operation 701A: Receiving first channel state information (CSI) sent by the first device. Operation 702A: Comparing the first CSI with a second CSI from the second device. Wherein, the first CSI is the original CSI input to the CSI encoder of the first device, and the second CSI is the recovered CSI output by the CSI decoder of the second device.

[0216] In some embodiments of this application, the artificial intelligence model monitoring method is executed on a first device, which is equipped with a Channel State Information (CSI) encoder applying an artificial intelligence model, and a second device is correspondingly equipped with a CSI decoder applying an artificial intelligence model. The method includes at least one of the following operations: sending a first Channel State Information (CSI) to the second device. This allows the first device to compare the first CSI with a second CSI from the second device. The first CSI is the original CSI input to the first device's CSI encoder, and the second CSI is the recovered CSI output from the second device's CSI decoder. By comparing the second CSI with the first CSI from the first device using the above technical solution, the reliability of the artificial intelligence model monitoring results can be improved, the signaling overhead of the artificial intelligence model monitoring can be reduced, and / or the air interface load can be reduced.

[0217] Specifically, the first device is, for example, the user equipment 120 or the base station 110 shown in Figure 2, and the second device is, for example, the corresponding base station 110 or user equipment 120 shown in Figure 2.

[0218] The fifth embodiment addresses the AI / ML-based CSI compression model monitoring mode on the network side. It employs a non-codebook approach to quantize and compress the raw CSI information (encoder input) obtained through CSI measurements. The quantized and compressed raw CSI information is then transmitted uplink to the gNB via the PUSCH channel, where the gNB performs SGCS / NMSE calculations and statistical operations. This fifth embodiment primarily addresses how to quantize and compress the raw CSI information input to the encoder, and the design of uplink signaling to ensure that the UE consistently understands the quantized and compressed CSI data and can correctly decode the raw CSI information.

[0219] In this fifth embodiment, the quantization and compression of the raw CSI information of the input encoder is performed using at least one of the following methods:

[0220] Method 1: Select a portion of the subband's recovered CSI for transmission.

[0221] In some embodiments of this application, the recovered CSI is a quantized and compressed recovered CSI based on a first quantization compression method, whereby the first quantization compression method involves transmitting the recovered CSI and / or recovered differential CSI of selected sub-bands. In some embodiments of this application, transmitting the recovered differential CSI includes transmitting the recovered CSI of a certain sub-band and the recovered differential CSI of other sub-bands. In some embodiments of this application, the difference between the recovered CSI sub-band differential CSI of the other sub-bands and the recovered CSI of the certain sub-band is obtained. In some embodiments of this application, the measured CSI sub-band differential information of all other sub-bands refers to: assuming there are N sub-bands, with indexes 0, 1, ..., N-1, transmitting the recovered CSI information of sub-band n, and subtracting the recovered CSI information of sub-bands 1, ..., n-1, n+1, ..., N-1 from the recovered CSI information of sub-band n to obtain the recovered CSI sub-band differential information of sub-bands 1, ..., n-1, n+1, ..., N-1. In some embodiments of this application, the first quantization compression method is to transmit the recovered CSI and / or the recovered differential CSI of the selected sub-bands based on rules or according to sub-band indicator instructions.

[0222] Specifically, since CSI information often exhibits high frequency domain correlation, meaning the precoding matrices of each sub-band have high similarity, it is unnecessary to calculate and statistically analyze w for each sub-band during model monitoring when calculating GSCS / NMSE monitoring performance metrics. The GSCS / NMSE allows for the selection of one or more representative subband precoding matrix coefficients for transmission and computation. Specific implementation methods include the following two approaches:

[0223] Alternative Solution 1: Rule-based.

[0224] In some embodiments of this application, selecting the recovery CSI and / or recovery differential CSI of the partial subbands for transmission based on the rules includes: selecting subbands whose channel quality conforms to preset rules. In some embodiments of this application, selecting the recovery CSI and / or recovery differential CSI of the partial subbands for transmission based on the rules includes: specifying that only the precoding matrix of the subband with the highest Channel Quality Indicator (CQI) is selected for calculation and statistics of model monitoring indicators. In some embodiments of this application, the preset rules are determined by one or more of the following methods: selecting all subbands exceeding a preset threshold; selecting N subbands exceeding the preset threshold; selecting K subbands with the highest channel quality, where K is greater than or equal to 1. The N subbands are subbands exceeding the preset threshold, and N can be greater than or equal to K. The indexes of the N subbands can be 0, 1, ..., N-1. The K subbands are the K subbands with the highest channel quality among the N subbands. In some embodiments of this application, when K equals 1, if there are multiple subbands with the same highest channel quality indicator value, only the subband corresponding to the first highest CQI in the reported CQI sequence sorted from left to right is selected, or only the subband corresponding to the first highest CQI in the high-order and low-order order of the reported CQI sequence is selected, or the subband is sorted by subband index, that is, only the subband with the largest or smallest subband index corresponding to the highest CQI in the reported CQI sequence is selected. In some embodiments of this application, the CQI sequence is included in the CSI report.

[0225] For example, it can be specified that only the precoding matrix of the subband with the highest CQI is selected for the calculation and statistics of the model monitoring index GSCS / NMSE. If there are multiple subbands with the highest CQI, only the subband corresponding to the first highest CQI that appears in the reported CQI sequence from left to right is selected.

[0226] Alternative 2: Based on the sub-band indicator.

[0227] In some embodiments of this application, transmission based on the selection of the recovery CSI and / or recovery differential CSI of the partial subbands indicated by the subband indicator includes: selecting the precoding matrix of the partial subbands for downlink transmission and representing it with a subband indicator bitmap. This refers to the UE automatically selecting the precoding matrix of the partial subbands for uplink transmission and representing it with a subband indicator bitmap. The subband indicator bitmap indicates the index of the selected subband. For example, if the BWP contains 10 subbands, and the UE selects the precoding matrices of the 1st and 3rd subbands for model monitoring, then the corresponding subband indicator bitmap is represented as "1010000000". The UE uploads the subband indicator bitmap and the selected partial subband precoding matrix data to the gNB. After receiving the information, the gNB parses the subband indicator bitmap and calculates the model monitoring index SGCS / NMSE based on the selected subbands. The specific basis and method for the UE to select subbands are implemented internally by the UE; it can be random selection or selection based on certain rules or algorithms.

[0228] Method 2: Select the precoding coefficients corresponding to a single polarization direction for transmission.

[0229] In some embodiments of this application, the recovered CSI is a quantized and compressed CSI based on a second quantization compression method, whereby the precoding coefficients corresponding to a single polarization direction are selected for transmission. In some embodiments of this application, selecting precoding coefficients corresponding to a single polarization direction for transmission includes: the precoding matrix of the single polarization direction carrying a polarization direction indicator to indicate the antenna polarization direction corresponding to the downlink transmission precoding matrix. In some embodiments of this application, the polarization direction indicator is represented by 1 bit, where a polarization direction indicator of 0 indicates an antenna port in the +45° polarization direction, and a polarization direction indicator of 1 indicates an antenna port in the -45° polarization direction. In some embodiments of this application, the polarization direction indicator is represented by 1 bit, where a polarization direction indicator of 1 indicates an antenna port in the +45° polarization direction, and a polarization direction indicator of 0 indicates an antenna port in the -45° polarization direction.

[0230] Specifically, the precoding matrix in the CSI information has a dimension of 2N1N2×v, meaning the number of rows corresponds to the number of CSI-RS ports, and the number of columns corresponds to the number of data layers. In some embodiments of this application, the upper N1N2 rows of the precoding matrix represent antenna ports in the +45° polarization direction, and the lower N1N2 rows represent antenna ports in the -45° polarization direction. In some embodiments of this application, the upper N1N2 rows of the precoding matrix represent antenna ports in the -45° polarization direction, and the lower N1N2 rows represent antenna ports in the +45° polarization direction. During model monitoring based on AI / ML CSI compression, the UE can select one of the precoding matrix coefficients for transmission in one polarization direction. After receiving the aforementioned CSI information, the gNB calculates w and in the corresponding polarization direction. The GSCS / NMSE is used. This halves the transmission overhead. Similar to method one, the precoding matrix for a single polarization direction transmitted by the UE uplink needs to carry a polarization direction indicator, indicating the antenna polarization direction corresponding to the uplink precoding matrix. In some embodiments of this application, the polarization direction indicator is represented by 1 bit, where 0 represents the antenna port in the +45° polarization direction (i.e., the upper N1N2 row of the complete precoding matrix), and 1 represents the antenna port in the -45° polarization direction (i.e., the lower N1N2 row of the complete precoding matrix). In some embodiments of this application, the polarization direction indicator is represented by 1 bit, where 1 represents the antenna port in the +45° polarization direction (i.e., the upper N1N2 row of the complete precoding matrix), and 0 represents the antenna port in the -45° polarization direction (i.e., the lower N1N2 row of the complete precoding matrix).

[0231] Method 3: Quantize the amplitude and phase coefficients of the precoding matrix.

[0232] In some embodiments of this application, the recovered CSI is a recovered CSI after quantization compression based on a third quantization compression method, whereby the third quantization compression method quantizes the amplitude coefficients and / or phase coefficients of the precoding matrix. In some embodiments of this application, the amplitude coefficients and / or the phase coefficients are quantized using 3-bit or 4-bit quantization. In some embodiments of this application, quantizing the amplitude coefficients and / or the phase coefficients of the precoding matrix includes amplitude quantization using amplitude coefficient quantization indicators and / or amplitude quantization using phase coefficient quantization indicators. In some embodiments of this application, the amplitude coefficient quantization indicator is represented by 1 bit; an amplitude coefficient quantization indicator of 0 indicates that the amplitude coefficient is quantized using 3 bits, and an amplitude coefficient quantization indicator of 1 indicates that the amplitude coefficient is quantized using 4 bits. In some embodiments of this application, the phase coefficient quantization indicator is represented by 1 bit; a phase coefficient quantization indicator of 0 indicates that the phase coefficient is quantized using 3 bits, and a phase coefficient quantization indicator of 1 indicates that the phase coefficient is quantized using 4 bits.

[0233] In some embodiments of this application, the artificial intelligence model monitoring method further includes: calculating the quantization error of the recovered CSI before and after the third quantization compression method. Specifically, if the compression method of quantizing the amplitude coefficients and phase coefficients of the precoding matrix of the recovered CSI information in method three is adopted, the resulting quantization error may affect the reliability of the calculation and statistics of the model monitoring performance indicators based on AI / ML CSI compression. Therefore, after quantizing the amplitude coefficients and phase coefficients of the precoding matrix, it is necessary to calculate the cosine similarity between the quantized precoding matrix and the precoding matrix recovered by the decoder output before quantization, and send this similarity to the UE.

[0234] Specifically, the coefficients of the elements of the unquantized precoding matrix are complex numbers, which can be represented as: The form is: where i, j represent the row and column numbers of the element coefficients in the precoding matrix, i = 1, 2, ..., 2N1N2, j = 1, ..., v, p ij θ represents the amplitude coefficient of the element. ij This represents the phase coefficient of the element. Generally, in CSI measurements, the obtained precoding matrix undergoes normalization, so the amplitude coefficient p... ij The range of values ​​for is 0 ≤ p ij ≤1, phase coefficient θ ij The range of values ​​for is 0 ≤ θ ijFor values ​​less than 2π, amplitude and phase coefficients are typically represented using floating-point numbers. Single-precision floating-point numbers require 32 bits, while double-precision floating-point numbers require 64 bits, resulting in significant overhead. Therefore, quantizing the amplitude and phase coefficients can greatly reduce data overhead.

[0235] In the fifth embodiment, the amplitude coefficient can employ either 3-bit or 4-bit quantization, and an amplitude coefficient quantization indicator is set to indicate which quantization method is selected. This indicator is 1 bit, and can be set to 0 for 3-bit quantization and 1 for 4-bit quantization. The UE determines which amplitude coefficient quantization method is selected and then notifies the gNB via the amplitude coefficient quantization indicator. The quantized value k of the amplitude coefficient and the amplitude coefficient p... ij The mapping relationship can be found in the amplitude quantization mapping table in Embodiment 1.

[0236] Similar to amplitude coefficient quantization, phase coefficient quantization typically employs either 3-bit or 4-bit quantization methods. The phase is quantized at equal intervals, and a phase coefficient quantization indicator is used to specify the selected quantization method. This indicator is 1 bit, and can be set to 0 for 3-bit quantization and 1 for 4-bit quantization. The specific phase coefficient quantization method selected is determined by the UE and then communicated to the UE via the phase coefficient quantization indicator. If 3-bit phase quantization is used, then the quantized value c of the phase coefficient... ij With phase coefficient θ ij The mapping formula is:

[0237] If 4-bit phase quantization is used, then the quantized value c of the phase coefficient is... ij With phase coefficient θ ij The mapping formula is:

[0238] If the compression method of quantizing the amplitude and phase coefficients of the precoding matrix for recovering CSI information (Method 3) is adopted, quantization errors will inevitably occur, thus affecting the reliability of the calculation and statistics of the model monitoring performance indicators based on AI / ML CSI compression. Therefore, after quantizing the amplitude and phase coefficients of the precoding matrix, it is necessary to calculate the cosine similarity between the quantized precoding matrix and the original precoding matrix, and upload this similarity to gNB as a basis for measuring the reliability of the GSCS / NMSE model monitoring performance indicators. The higher the cosine similarity, the smaller the quantization error generated after quantization, and the more reliable the calculation and statistics of the model monitoring performance indicators.

[0239] Suppose that the precoding matrix w from the original precoding matrix w of the input encoder, after amplitude and phase quantization, is transformed back to the precoding matrix w. q Then, the cosine similarity between them is calculated as follows:

[0240] Cosine similarity SGCS q SGCS characterizes the magnitude of the quantization error after amplitude and phase quantization of w. q The larger the value, the smaller the quantization error. (Using w) q The greater the statistical reliability of the calculated performance metric SGCS / NMSE for model monitoring, the better. (SGCS is a cosine similarity metric.) q The quantized values ​​need to be sent to the UE so that the UE can measure the reliability of the model monitoring performance index calculation results and determine whether the performance index results of this model monitoring are usable.

[0241] Cosine similarity SGCS q The specific quantization method is the same as the cosine similarity SGCS in the first embodiment. q Quantification methods.

[0242] In summary, network-side CSI compression model monitoring based on non-codebook transmission requires the UE to employ at least one of the methods 1-3 to quantize and compress the raw CSI information input to the encoder, and then transmit the quantized and compressed CSI data uplink to the gNB. To ensure that the gNB can accurately parse the quantized and compressed CSI data, the entire uplink CSI data is transmitted using a first-part and second-part mode. The second part corresponds to the quantized and compressed CSI data of the input encoder, while the first part is a set of various indicators for quantizing and compressing the raw CSI information. The bit format of the first part is described below.

[0243] In some embodiments of this application, the recovered CSI includes first information (or a first part) and second information (or a second part), wherein the first information is a set of indicators for quantizing and compressing the recovered CSI, and the second information is CSI data for quantizing and compressing the recovered CSI. In some embodiments of this application, the first information is received through static indicators, periodic configuration, semi-static indicators, or dynamic indicators. In some embodiments of this application, when the recovered CSI and / or the recovered differential CSI of the selected sub-band are transmitted based on rules, the data composition of the first information from the most significant bit (MSB) to the least significant bit (LSB) represents: quantization compression mode indicator: corresponding to whether the first quantization compression method, the second quantization compression method, and the third quantization compression method are used to quantize and compress the CSI information, wherein the first quantization compression method is to select the recovered CSI and / or the recovered differential CSI of the selected sub-band for transmission, the second quantization compression method is to select the precoding coefficients corresponding to a single polarization direction for transmission, and the third quantization compression method is to quantize the amplitude coefficients and / or phase coefficients of the precoding matrix; polarization direction selection indicator: indicating the antenna polarization direction corresponding to the downlink transmission precoding matrix; amplitude coefficient quantization indicator: indicating the quantization method of the amplitude coefficients of the precoding matrix elements; phase coefficient quantization indicator: indicating the quantization method of the phase coefficients of the precoding matrix elements; quantization compression similarity indicator: indicating the quantized value of the cosine similarity after quantization compression; second information data size indicator: indicating the number of data bits of the second information.

[0244] Case 1: Subband selection for CSI recovery uses a rule-based selection method. The first part of the data, from MSB to LSB, represents:

[0245] Quantization compression mode indicator (3 bits): corresponds to whether to use mode 1, mode 2 and mode 3 to quantize and compress CSI information, where "0" means not to use and "1" means to use.

[0246] Polarization direction selection indicator (1 bit): Indicates the antenna polarization direction corresponding to the precoding matrix of the downlink transmission. If mode 2 is not used, this bit is meaningless.

[0247] Amplitude coefficient quantization indicator (1 bit): Indicates the quantization method of the amplitude coefficient of the precoding matrix element. If mode 3 is not used, this bit is meaningless.

[0248] Phase coefficient quantization indicator (1 bit): Indicates the quantization method of the phase coefficient of the precoded matrix element. If mode 3 is not used, this bit is meaningless.

[0249] Quantization Compression Similarity Indicator (4 bits): Indicates the cosine similarity (SGCS) after quantization compression. q The quantization value is meaningless if method 3 is not used.

[0250] The data size of the second part (X-bit): The number of data bits in the second part. Similar to the first embodiment, it can be indicated explicitly or implicitly.

[0251] In some embodiments of this application, when the recovery CSI and / or recovery differential CSI of a portion of the subband is selected for transmission according to the subband indicator, the data composition of the first information from the most significant bit (MSB) to the least significant bit (LSB) represents: quantization compression mode indicator: corresponding to whether the first quantization compression method, the second quantization compression method, and the third quantization compression method are used to quantize and compress the CSI information, respectively. The first quantization compression method involves selecting the recovery CSI and / or recovery differential CSI of a portion of the subband for transmission, and the second quantization compression method involves selecting the precoding coefficients corresponding to a single polarization direction for transmission. The third quantization compression method quantizes the amplitude coefficients and / or phase coefficients of the precoding matrix; the polarization direction selection indicator indicates the antenna polarization direction corresponding to the downlink transmission precoding matrix; the amplitude coefficient quantization indicator indicates the quantization method of the amplitude coefficients of the precoding matrix elements; the phase coefficient quantization indicator indicates the quantization method of the phase coefficients of the precoding matrix elements; the quantization compression similarity indicator indicates the quantized value of the cosine similarity after quantization compression; the second information data size indicator indicates the number of data bits of the second information; and the subband indicator indicates the selected subband in a bitmap format.

[0252] Case 2: The subband selection for CSI recovery uses a selection method based on subband indicator. The first part of the data, from MSB to LSB, represents:

[0253] Quantization compression mode indicator (3 bits): corresponds to whether to use mode 1, mode 2 and mode 3 to quantize and compress CSI information, where "0" means not to use and "1" means to use.

[0254] Subband Indicator (N3bit): The subband indicator bitmap indicates the index of the selected subband.

[0255] Polarization direction selection indicator (1 bit): Indicates the antenna polarization direction corresponding to the precoding matrix of the downlink transmission. If mode 2 is not used, this bit is meaningless.

[0256] Amplitude coefficient quantization indicator (1 bit): Indicates the quantization method of the amplitude coefficient of the precoding matrix element. If mode 3 is not used, this bit is meaningless.

[0257] Phase coefficient quantization indicator (1 bit): Indicates the quantization method of the phase coefficient of the precoded matrix element. If mode 3 is not used, this bit is meaningless.

[0258] Quantization Compression Similarity Indicator (4 bits): Indicates the cosine similarity (SGCS) after quantization compression. q The quantization value is meaningless if method 3 is not used.

[0259] The data size of the second part (X-bit): The number of data bits in the second part. Similar to Example 1, it can be indicated explicitly or implicitly.

[0260] Subband indicator (N3bit): Indicates the selected subband to be transmitted using a bitmap.

[0261] The general process of the fifth embodiment is shown in Figure 7B. The network-side model monitoring operation steps include at least one of the following steps: Step 1: The gNB sends a CSI-RS reference signal to the UE. The UE measures the actual CSI information (i.e., the raw CSI information) based on the reference signal, inputs it into the encoder for CSI compression, generates compressed CSI feedback information, and reports the compressed CSI feedback information and the actual CSI information together to the gNB. Step 2: After receiving the above information, the gNB inputs the compressed CSI feedback information into the decoder for decompression, recovers the CSI information, and calculates the model monitoring index SGCS with the reported actual CSI information.

[0262] As shown above, the number of bits in the first part of the data is fixed. The UE can accurately parse the information in the first part using blind detection, and further accurately parse the data in the second part. The first part of the data and the second part of the data can be transmitted separately or together. The first part of the data and the second part of the data can be transmitted via Uplink Control Information (UCI) signaling; or the first part can be transmitted via UCI signaling and the second part via UL MAC-CE; or the first part can be transmitted via UCI signaling and the second part via uplink data transmission in PUSCH.

[0263] Sixth Implementation: Monitoring of Network-Side CSI Compression Model Based on Codebook Transmission:

[0264] The sixth embodiment can be implemented independently. In some embodiments of this application, the sixth embodiment can also be combined with other embodiments, for example, the sixth embodiment can be implemented in combination with the first embodiment, the second embodiment, the third embodiment, the fourth embodiment, and / or the fifth embodiment. These embodiments can be implemented sequentially or in parallel, and this application is not limited thereto.

[0265] Figure 8A is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application. As shown in Figure 8A, the artificial intelligence model monitoring method is executed on a second device. The second device is equipped with a Channel State Information (CSI) decoder that applies an artificial intelligence model, and the first device is equipped with a CSI encoder that applies an artificial intelligence model. The method includes at least one of the following operations: Operation 801A: Receiving a first precoding matrix indication (PMI) sent by the first device. Operation 802A: Comparing the first PMI with the first PMI of the second device. Wherein, the second PMI is the corresponding PMI of the recovered CSI output by the CSI decoder of the second device, and the first PMI is the corresponding PMI of the original CSI input by the CSI encoder of the first device.

[0266] In some embodiments of this application, the artificial intelligence model monitoring method is executed on a first device, the first device being equipped with a Channel State Information (CSI) encoder applying an artificial intelligence model, and a second device correspondingly equipped with a CSI decoder applying an artificial intelligence model. The method includes at least one of the following operations: sending a first precoding matrix indication (PMI) to the second device. This allows the first device to compare the first PMI with a second PMI of the second device. The second PMI is the corresponding PMI of the recovered CSI output by the second device's CSI decoder, and the first PMI is the corresponding PMI of the original CSI input by the first device's CSI encoder.

[0267] By comparing the second PMI with the first PMI of the first device using the above technical solution, the reliability of the AI ​​model monitoring results can be improved, the signaling overhead of the AI ​​model monitoring can be reduced, and / or the air interface load can be reduced.

[0268] Specifically, the first device is, for example, the user equipment 120 or the base station 110 shown in Figure 2, and the second device is, for example, the corresponding base station 110 or user equipment 120 shown in Figure 2.

[0269] In some embodiments of this application, the corresponding PMI for the recovered CSI includes first information (or a first part) and second information (or a second part). The first information includes configuration parameters related to codebook compression, and the second information is CSI data for codebook-based quantization compression of the recovered CSI. In some embodiments of this application, the configuration parameters related to codebook compression include codebook type, the number of spatial basis selections, the number of frequency basis selections, and the maximum number of non-zero coefficients in the weighting amplitude coefficients of the sub-band components. In some embodiments of this application, the corresponding PMI for the recovered CSI corresponds to a parameter combination of the codebook type. In some embodiments of this application, the parameters of the corresponding PMI for the recovered CSI are indicated in the first information. In some embodiments of this application, the first information is received through static indication, periodic configuration, semi-static indication, or dynamic indication.

[0270] The sixth embodiment is for the model monitoring mode based on AI / ML CSI compression on the network side. It adopts a codebook-based approach to compress and encode the original CSI information of the input encoder into PMI. Then, the PMI and the compressed CSI are transmitted uplink to the gNB through the PUCCH / PUSCH channel. The gNB performs SGCS / NMSE calculation and statistical operations.

[0271] The codebook reporting type of the PMI is configured by the higher-layer parameter `codebookType`. Depending on the specific codebook type, related parameters (e.g., the number of spatial basis selections `L` in the version 15 Type II codebook, and the parameter combination index in the version 16e Type II codebook and version 17 Fe Type II codebook) are also configured by higher-layer parameters. The UE must adhere to the parameters configured by the gNB during PMI generation. After generating the PMI, the UE transmits the PMI along with the compressed CSI output by the encoder to the gNB via PUCCH / PUSCH information. CSI reporting uses a three-part method:

[0272] Part 1: Includes the number of non-zero wideband amplitude coefficients for each layer of RI, CRI, CQI, and Type I ICSI (applicable to Type II codebooks) and the number of bits for compressed CSI (Part 3).

[0273] Part Two: Includes PMI and subband CQI when RI is greater than 4 (applicable to type I codebook).

[0274] Part Three: Contains compressed CSI.

[0275] The general process for the fifth and sixth embodiments is shown in Figure 7B. The network-side model monitoring operation steps include at least one of the following steps: Step 1: The gNB sends a CSI-RS reference signal to the UE. The UE measures the actual CSI information (i.e., the raw CSI information) based on the reference signal, inputs it into the encoder for CSI compression, generates compressed CSI feedback information, and reports the compressed CSI feedback information and the actual CSI information together to the gNB. Step 2: After receiving the above information, the gNB inputs the compressed CSI feedback information into the decoder for decompression, recovers the CSI information, and calculates the model monitoring index SGCS with the reported actual CSI information.

[0276] Figure 8B is a flowchart illustrating the artificial intelligence model monitoring method provided in an embodiment of this application. As shown in Figure 8B, the artificial intelligence model monitoring method is executed on a second device. The second device is equipped with a Channel State Information (CSI) decoder that applies an artificial intelligence model, and the first device is equipped with a CSI encoder that applies an artificial intelligence model. The method includes at least one of the following operations: Operation 801B: Determine the corresponding second downlink channel quality information based on the recovered CSI. Operation 802B: Receive the first downlink channel quality information sent by the first device, wherein the first downlink channel quality information is determined based on the original CSI. Operation 803B: Compare the model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device. In some embodiments of this application, the artificial intelligence model monitoring method is executed on a first device, which is equipped with a Channel State Information (CSI) encoder applying an artificial intelligence model, and a second device is equipped with a CSI decoder applying an artificial intelligence model. The method includes at least one of the following operations: sending first downlink channel quality information to the second device. The first downlink channel quality information is determined based on the original CSI. This allows the second device to compare model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information of the second device. The second downlink channel quality information is determined based on recovered CSI. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device.

[0277] In some embodiments of this application, the artificial intelligence model monitoring method is executed on a first device, the first device being equipped with a channel state information (CSI) encoder applying an artificial intelligence model, and a second device being equipped with a corresponding CSI decoder applying an artificial intelligence model. The method includes at least one of the following operations: sending first downlink channel quality information to the second device. The first downlink channel quality information is determined based on the original CSI; comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information of the second device. The second downlink channel quality information is determined based on recovered CSI. The recovered CSI is the recovered CSI output by the CSI decoder of the second device, and the original CSI is the original CSI input by the CSI encoder of the first device.

[0278] By employing the above technical solution, the model monitoring performance indicators are compared based on the first downlink channel quality information and the second downlink channel quality information. This improves the reliability of the AI ​​model monitoring results, reduces the signaling overhead of AI model monitoring, and / or lowers the air interface load.

[0279] Specifically, the first device is, for example, the user equipment 120 or the base station 110 shown in Figure 2, and the second device is, for example, the corresponding base station 110 or user equipment 120 shown in Figure 2.

[0280] In some embodiments of this application, the first downlink channel quality information and the second downlink channel quality information include Signal-to-Interference-plus-Noise Ratio (SINR), Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), Received Signal Strength Indication (RSSI), or Channel Quality Indicator (CQI). In some embodiments of this application, the difference between the first downlink channel quality information and the second downlink channel quality information is greater than a preset threshold. In some embodiments of this application, comparing the model monitoring performance index based on the first downlink channel quality information and the second downlink channel quality information includes comparing the similarity between the first downlink channel quality information and the second downlink channel quality information. In some embodiments of this application, comparing the model monitoring performance index between the first downlink channel quality information and the second downlink channel quality information includes comparing the cosine similarity between the first downlink channel quality information and the cosine similarity between the original CSI and the recovered CSI. In some embodiments of this application, the cosine similarity and / or normalized mean square error between the first downlink channel quality information and the second downlink channel quality information are used as the model monitoring performance index.

[0281] Figure 9 is a schematic structural diagram of a wireless communication device 700 provided in an embodiment of this application. This wireless communication device can be a user equipment, a base station, or a network element. The wireless communication device 700 shown in Figure 9 includes a processor 710, which can call and run computer programs from memory to implement the methods in the embodiments of this application.

[0282] Optionally, as shown in FIG10, the wireless communication device 700 may further include a memory 720. The processor 710 can retrieve and run computer programs from the memory 720 to implement the methods in the embodiments of this application. The memory 720 may be a separate device independent of the processor 710, or it may be integrated into the processor 710.

[0283] Optionally, as shown in Figure 9, the wireless communication device 700 may further include a transceiver 730. The processor 710 can control the transceiver 730 to communicate with other devices. Specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 730 may include a transmitter and a receiver. The transceiver 730 may further include an antenna, and the number of antennas may be one or more.

[0284] Optionally, the wireless communication device 700 may specifically be a base station in the embodiments of this application, and the wireless communication device 700 may implement the corresponding processes implemented by the base station in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0285] Optionally, the wireless communication device 700 may specifically be a mobile user equipment / user equipment in the embodiments of this application, and the wireless communication device 700 may implement the corresponding processes implemented by the mobile user equipment / user equipment in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0286] Optionally, the wireless communication device 700 may specifically be an LMF entity in the embodiments of this application, and the wireless communication device 700 may implement the corresponding processes implemented by the network element in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0287] Figure 10 is a schematic structural diagram of a chip according to an embodiment of this application. The chip 800 shown in Figure 10 includes a processor 810, which can call and run computer programs from memory to implement the methods in the embodiments of this application.

[0288] Optionally, as shown in FIG10, chip 800 may further include memory 820. Processor 810 can call and run computer programs from memory 820 to implement the methods in the embodiments of this application. Memory 820 may be a separate device independent of processor 810, or it may be integrated into processor 810.

[0289] Optionally, the chip 800 may also include an input interface 830. The processor 910 can control the input interface 830 to communicate with other devices or chips; specifically, it can acquire information or data sent by other devices or chips.

[0290] Optionally, the chip 800 may also include an output interface 840. The processor 810 can control the output interface 840 to communicate with other devices or chips, specifically, to output information or data to other devices or chips.

[0291] Optionally, the chip can be applied to the base station in the embodiments of this application, and the chip can implement the corresponding processes implemented by the base station in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0292] Optionally, the chip can be applied to the mobile user equipment / user equipment in the embodiments of this application, and the chip can implement the corresponding processes implemented by the mobile user equipment / user equipment in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0293] Optionally, the chip can be applied to the network element in the embodiments of this application, and the chip can implement the corresponding processes implemented by the mobile network element in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0294] Figure 11 is a schematic block diagram of a wireless communication system 100 provided in an embodiment of this application. As shown in Figure 11, the communication system 100 includes a user equipment 120 and a base station 110. The user equipment 120 can be used to implement the corresponding functions implemented by the user equipment in the above method, and the base station 110 can be used to implement the corresponding functions implemented by the base station in the above method. For simplicity, these will not be described in detail here.

[0295] It should be understood that the processor in the embodiments of this application may be an integrated circuit chip with signal processing capabilities. During implementation, the steps of the above method embodiments can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. It is understood that the memory in the embodiments of this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory. Embodiments of this application also provide a computer-readable storage medium for storing computer programs.

[0296] Optionally, the computer-readable storage medium can be applied to the base station in the embodiments of this application, and the computer program causes the computer to execute the corresponding processes implemented by the base station in the various methods of the embodiments of this application. For simplicity, further details are omitted here. Optionally, the computer-readable storage medium can be applied to the mobile user equipment / user equipment in the embodiments of this application, and the computer program causes the computer to execute the corresponding processes implemented by the mobile user equipment / user equipment in the various methods of the embodiments of this application. For simplicity, further details are omitted here.

[0297] This application also provides a computer program product, including computer program instructions.

[0298] Optionally, the computer program product can be applied to the base station in the embodiments of this application, and the computer program instructions cause the computer to execute the corresponding processes implemented by the base station in the various methods of the embodiments of this application. For simplicity, further details are omitted here. Optionally, the computer program product can be applied to the mobile user equipment / user equipment in the embodiments of this application, and the computer program instructions cause the computer to execute the corresponding processes implemented by the mobile user equipment / user equipment in the various methods of the embodiments of this application. For simplicity, further details are omitted here.

[0299] This application also provides a computer program.

[0300] Optionally, the computer program can be applied to the base station in the embodiments of this application. When the computer program runs on a computer, it causes the computer to execute the corresponding processes implemented by the base station in the various methods of the embodiments of this application. For simplicity, these will not be described in detail here. Optionally, the computer program can be applied to the mobile user equipment / user equipment in the embodiments of this application. When the computer program runs on a computer, it causes the computer to execute the corresponding processes implemented by the mobile user equipment / user equipment in the various methods of the embodiments of this application. For simplicity, these will not be described in detail here.

[0301] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0302] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An artificial intelligence model monitoring method, executed on a first device, the first device being deployed with a channel state information (CSI) encoder applying an artificial intelligence model, a second device being deployed with a CSI decoder applying an artificial intelligence model, the method comprising: receiving second CSI sent by the second device; and comparing the second CSI with first CSI of the first device; wherein the second CSI is recovered CSI output by the CSI decoder of the second device, and the first CSI is original CSI input to the CSI encoder of the first device; the recovered CSI is quantized and compressed recovered CSI based on a first quantization compression manner, the first quantization compression manner being to select partial subbands of recovered CSI and / or recovered differential CSI for transmission; selecting recovered differential CSI for transmission comprises transmitting recovered CSI of a certain subband and recovered differential CSI of other subbands; the recovered differential CSI of the other subbands is obtained respectively according to differences between the recovered CSI of the other subbands and the recovered CSI of the certain subband; the first quantization compression manner is to select the partial subbands of recovered CSI and / or the recovered differential CSI for transmission based on a rule or according to a subband indicator; selecting the partial subbands of recovered CSI and / or the recovered differential CSI for transmission based on the rule comprises selecting subbands with channel quality meeting a preset rule; the preset rule is determined by one or more of the following manners: selecting all subbands exceeding a preset threshold; selecting N subbands exceeding the preset threshold; selecting K subbands with highest channel quality, K being greater than or equal to 1; when K is equal to 1, if there are multiple subbands with the same highest channel quality indicator (CQI) value, only taking a subband corresponding to a first highest CQI in a left-to-right order in a reported CQI sequence, or only taking a subband corresponding to a first highest CQI in a high-to-low order in the reported CQI sequence, or only taking a subband with a maximum or minimum subband index corresponding to a highest CQI in the reported CQI sequence; the CQI sequence is contained in a CSI report; selecting the partial subbands of recovered CSI and / or the recovered differential CSI for transmission according to the subband indicator comprises selecting precoding matrices of partial subbands of downlink transmission and representing the precoding matrices by a subband indicator bitmap; the recovered CSI is quantized and compressed recovered CSI based on a second quantization compression manner, the second quantization compression manner being to select precoding coefficients corresponding to a single polarization direction for transmission; selecting precoding coefficients corresponding to a single polarization direction for transmission comprises a precoding matrix of the single polarization direction carrying a polarization direction indicator indicating an antenna polarization direction corresponding to a precoding matrix of downlink transmission; the polarization direction indicator is represented by one bit, the polarization direction indicator being 0 indicating an antenna port of a +45° polarization direction, and the polarization direction indicator being 1 indicating an antenna port of a -45° polarization direction. ​ ​ ​ ​ ​ 2.The artificial intelligence model monitoring method of claim 1, wherein, ​ 3.The artificial intelligence model monitoring method of claim 2, wherein, ​ 4.The artificial intelligence model monitoring method of claim 3, wherein, ​ 5.The artificial intelligence model monitoring method of claim 2, wherein, ​ 6.The artificial intelligence model monitoring method of claim 5, wherein, ​ 7.The artificial intelligence model monitoring method of claim 6, wherein, ​ ​ ​ ​ 8.The artificial intelligence model monitoring method of claim 7, wherein, ​ 9.The artificial intelligence model monitoring method of claim 8, wherein, ​ 10.The artificial intelligence model monitoring method of claim 5, wherein, ​ 11.The AI model monitoring method of claim 1, wherein, ​ 12.The artificial intelligence model monitoring method of claim 11, wherein, ​ 13.The artificial intelligence model monitoring method of claim 12, wherein, ​ 14.The AI model monitoring method of claim 1, wherein, The recovered CSI is quantization-compressed recovered CSI based on a third quantization compression mode, and the third quantization compression mode is quantization of amplitude coefficients and / or phase coefficients of a precoding matrix. 15.The artificial intelligence model monitoring method of claim 14, wherein, The amplitude coefficients and / or the phase coefficients are quantized by 3-bit quantization or 4-bit quantization. 16.The AI model monitoring method of claim 14, wherein, The quantization of the amplitude coefficients and / or the phase coefficients of the precoding matrix includes indicating the amplitude quantization mode of the amplitude coefficients by an amplitude coefficient quantization indicator and / or indicating the amplitude quantization mode of the phase coefficients by the phase coefficient quantization indicator. 17.The artificial intelligence model monitoring method of claim 16, wherein, The amplitude coefficient quantization indicator is 1-bit, and the amplitude coefficient quantization indicator is 0 to indicate that the amplitude coefficients are quantized by 3-bit quantization, and the amplitude coefficient quantization indicator is 1 to indicate that the amplitude coefficients are quantized by 4-bit quantization. 18.The AI model monitoring method of claim 16, wherein, The phase coefficient quantization indicator is 1-bit, and the phase coefficient quantization indicator is 0 to indicate that the phase coefficients are quantized by 3-bit quantization, and the phase coefficient quantization indicator is 1 to indicate that the phase coefficients are quantized by 4-bit quantization.

19. The artificial intelligence model monitoring method of claim 14, further comprising: receiving the quantization error of the recovered CSI before and after the third quantization compression mode sent by the second device. 20.The AI model monitoring method of claim 1, wherein, The recovered CSI includes first information and second information, the first information is a set of indicators for quantization compression of the recovered CSI, and the second information is CSI data for quantization compression of the recovered CSI. 21.The artificial intelligence model monitoring method of claim 20, wherein, The first information is received by static indication, periodic configuration, semi-static indication, or dynamic indication. 22.The AI model monitoring method of claim 20, wherein, When the recovered CSI of the part of the subbands and / or the recovered differential CSI is selected for transmission based on the rule, the data composition of the first information from the most significant bit MSB to the least significant bit LSB respectively represents: a quantization compression mode indicator corresponding to whether the CSI information is quantization-compressed by a first quantization compression mode, a second quantization compression mode, and a third quantization compression mode, wherein the first quantization compression mode is to select the recovered CSI of the part of the subbands and / or the recovered differential CSI for transmission, the second quantization compression mode is to select the precoding coefficients corresponding to a single polarization direction for transmission, and the third quantization compression mode is to quantize amplitude coefficients and / or phase coefficients of a precoding matrix; a polarization direction selection indicator indicating the antenna polarization direction corresponding to the precoding matrix of the downlink transmission; an amplitude coefficient quantization indicator indicating the quantization mode of the amplitude coefficients of the precoding matrix elements; a phase coefficient quantization indicator indicating the quantization mode of the phase coefficients of the precoding matrix elements; a quantization compression similarity indicator indicating the quantized value of the cosine similarity after quantization compression; a second information data size indicator indicating the number of data bits of the second information. 23.The AI model monitoring method of claim 20, wherein, When the recovered CSI of the part of the subbands and / or the recovered differential CSI is selected for transmission according to the subband indicator, the data composition of the first information from the most significant bit MSB to the least significant bit LSB respectively represents: a quantization compression mode indicator, which respectively corresponds to whether the CSI information is quantized and compressed by a first quantization compression mode, a second quantization compression mode and a third quantization compression mode, wherein the first quantization compression mode is to select a part of the recovered CSI and / or the recovered differential CSI for transmission, the second quantization compression mode is to select the precoding coefficients corresponding to a single polarization direction for transmission, and the third quantization compression mode is to quantize the amplitude coefficients and / or the phase coefficients of the precoding matrix; a polarization direction selection indicator, which indicates the antenna polarization direction corresponding to the precoding matrix of the downlink transmission; an amplitude coefficient quantization indicator, which indicates the quantization mode of the amplitude coefficients of the precoding matrix elements; a phase coefficient quantization indicator, which indicates the quantization mode of the phase coefficients of the precoding matrix elements; a quantization compression similarity indicator, which indicates the quantized value of the cosine similarity after quantization and compression; a second information data size indicator, which indicates the data bit number of the second information by using the second information data size indicator; a subband indicator, which indicates the selected subband in the form of a bitmap.

24. An artificial intelligence model monitoring method, executed on a first device, wherein the first device is deployed with a channel state information (CSI) encoder applying an artificial intelligence model, and a second device is correspondingly deployed with a CSI decoder applying an artificial intelligence model, and the method comprises: receiving a second precoding matrix indicator (PMI) sent by the second device; comparing the second PMI with a first PMI of the first device; wherein the second PMI is a corresponding PMI of a recovered CSI output by the CSI decoder of the second device, and the first PMI is a corresponding PMI of an original CSI input by the CSI encoder of the first device. the corresponding PMI of the recovered CSI comprises first information and second information, the first information comprises codebook compression related configuration parameters, and the second information is CSI data obtained by codebook based quantization and compression of the recovered CSI. the codebook compression related configuration parameters comprise a codebook type, a selected number of spatial domain bases, a selected number of frequency domain bases, and a maximum number of non-zero coefficients of weighted amplitude coefficients of subband components.

25. The artificial intelligence model monitoring method of claim 24, wherein, the corresponding PMI of the recovered CSI corresponds to a parameter combination of the codebook type.

26. The artificial intelligence model monitoring method of claim 54, wherein, parameters of the corresponding PMI of the recovered CSI are indicated in the first information.

27. The artificial intelligence model monitoring method of claim 26, wherein, the first information is received by static indication, periodic configuration, semi-static indication or dynamic indication.

28. The artificial intelligence model monitoring method of claim 26, wherein, 30. An artificial intelligence model monitoring method, executed on a first device, wherein the first device is deployed with a channel state information (CSI) encoder applying an artificial intelligence model, and a second device is correspondingly deployed with a CSI decoder applying an artificial intelligence model, and the method comprises:

29. The artificial intelligence model monitoring method of claim 25, wherein, obtaining an original CSI by CSI measurement, and determining a corresponding first downlink channel quality information according to the original CSI; receiving second downlink channel quality information sent by the second device, wherein the second downlink channel quality information is determined based on a recovered CSI; comparing a model monitoring performance indicator based on the first downlink channel quality information and the second downlink channel quality information; wherein ​ ​ ​ The recovered CSI is a recovered CSI output by a CSI decoder of the second device, and the original CSI is an original CSI input by a CSI encoder of the first device.

31. The artificial intelligence model monitoring method of claim 30, wherein, The first and second downlink channel quality information includes signal-to-interference-plus-noise ratio (SINR), reference signal received power (RSRP), reference signal received quality (RSRQ), received signal strength indicator (RSSI), or channel quality indicator (CQI).

32. The artificial intelligence model monitoring method of claim 30, wherein, The difference between the first and second downlink channel quality information is greater than a preset threshold.

33. The artificial intelligence model monitoring method of claim 30, wherein, Comparing the model monitoring performance indicators based on the first and second downlink channel quality information includes: Comparing the similarity of the first and second downlink channel quality information.

34. The artificial intelligence model monitoring method of claim 30, wherein, Comparing the model monitoring performance indicators based on the first and second downlink channel quality information includes: The positive correlation between the cosine similarity of the first and second downlink channel quality information and the cosine similarity between the original and recovered CSI.

35. The artificial intelligence model monitoring method of claim 30, wherein, The cosine similarity and / or normalized mean square error between the first and second downlink channel quality information as the model monitoring performance indicators.

36. An artificial intelligence model monitoring method, executed on a second device, the second device being deployed with a channel state information (CSI) decoder applying an artificial intelligence model, a first device corresponding to being deployed with a CSI encoder applying an artificial intelligence model, the method comprising: receiving first channel state information (CSI) sent by the first device; and comparing the first CSI with second CSI of the second device; wherein the first CSI is original CSI input by a CSI encoder of the first device, and the second CSI is recovered CSI output by a CSI decoder of the second device. The recovered CSI is quantized and compressed based on a first quantization compression manner, and the first quantization compression manner is to select partial subband recovered CSI and / or recovered differential CSI for transmission.

37. The artificial intelligence model monitoring method of claim 36, wherein, Selecting recovered differential CSI for transmission includes transmitting recovered CSI of a certain subband and recovered differential CSI of other subbands.

38. The artificial intelligence model monitoring method of claim 37, wherein, The measured CSI subband differential CSI of the other subbands is obtained respectively according to the difference between the recovered CSI of the other subbands and the recovered CSI of the certain subband.

39. The artificial intelligence model monitoring method of claim 38, wherein, The first quantization compression manner is to select the partial subband recovered CSI and / or the recovered differential CSI for transmission based on a rule or according to a subband indicator.

40. The artificial intelligence model monitoring method of claim 37, wherein, Selecting the partial subband recovered CSI and / or the recovered differential CSI for transmission based on the rule includes selecting subbands with channel quality meeting a preset rule.

41. The artificial intelligence model monitoring method of claim 40, wherein, The preset rule is determined by one or more of the following ways:

42. The artificial intelligence model monitoring method of claim 41, wherein, selecting all subbands exceeding a preset threshold; selecting N subbands exceeding the preset threshold; selecting K subbands with the highest channel quality, K being greater than or equal to 1. ​ 43. The artificial intelligence model monitoring method of claim 42, wherein, When K is equal to 1, if there are multiple subbands with the same highest channel quality indicator (CQI) value, only the subband corresponding to the first highest CQI in the reported CQI sequence from left to right ordering, or only the subband corresponding to the first highest CQI in the reported CQI sequence from high to low ordering, or only the subband with the largest or smallest subband index corresponding to the highest CQI in the reported CQI sequence is selected.

44. The artificial intelligence model monitoring method of claim 43, wherein, The CQI sequence is included in a CSI report.

45. The artificial intelligence model monitoring method of claim 40, wherein, The transmitting the recovered CSI and / or the recovered differential CSI of the selected part of the subbands based on the selection of the part of the subbands indicated by the subband indicator comprises selecting a precoding matrix of the part of the subbands of the downlink transmission and representing the precoding matrix with a subband indication bitmap.

46. The artificial intelligence model monitoring method of claim 36, wherein, The recovered CSI is quantized and compressed based on a second quantization and compression manner, and the second quantization and compression manner is transmitting precoding coefficients corresponding to a single polarization direction.

47. The artificial intelligence model monitoring method of claim 46, wherein, The transmitting precoding coefficients corresponding to a single polarization direction comprises that a precoding matrix of the single polarization direction carries a polarization direction indicator to indicate an antenna polarization direction corresponding to a precoding matrix of the downlink transmission.

48. The artificial intelligence model monitoring method of claim 47, wherein, The polarization direction indicator is represented by 1 bit, and the polarization direction indicator is 0 to represent an antenna port of a +45° polarization direction, and the polarization direction indicator is 1 to represent an antenna port of a -45° polarization direction.

49. The artificial intelligence model monitoring method of claim 36, wherein, The recovered CSI is quantized and compressed based on a third quantization and compression manner, and the third quantization and compression manner is quantizing amplitude coefficients and / or phase coefficients of the precoding matrix.

50. The artificial intelligence model monitoring method of claim 49, wherein, The amplitude coefficients and / or the phase coefficients are quantized by 3 bits or 4 bits.

51. The artificial intelligence model monitoring method of claim 49, wherein, The quantizing the amplitude coefficients and / or the phase coefficients of the precoding matrix comprises indicating an amplitude quantization manner of the amplitude coefficients by an amplitude coefficient quantization indicator and / or indicating an amplitude quantization manner of the phase coefficients by a phase coefficient quantization indicator.

52. The artificial intelligence model monitoring method of claim 51, wherein, The amplitude coefficient quantization indicator is represented by 1 bit, and the amplitude coefficient quantization indicator is 0 to represent that the amplitude coefficients are quantized by 3 bits, and the amplitude coefficient quantization indicator is 1 to represent that the amplitude coefficients are quantized by 4 bits.

53. The artificial intelligence model monitoring method of claim 51, wherein, The phase coefficient quantization indicator is represented by 1 bit, and the phase coefficient quantization indicator is 0 to represent that the phase coefficients are quantized by 3 bits, and the phase coefficient quantization indicator is 1 to represent that the phase coefficients are quantized by 4 bits.

54. The artificial intelligence model monitoring method of claim 49, further comprising: calculating a quantization error of the recovered CSI before and after the third quantization and compression manner.

55. The artificial intelligence model monitoring method of claim 36, wherein, The recovered CSI comprises first information and second information, the first information is a set of indicators of quantization and compression of the recovered CSI, and the second information is CSI data of quantization and compression of the recovered CSI.

56. The artificial intelligence model monitoring method of claim 55, wherein, The first information is received by static indication, periodic configuration, semi-static indication, or dynamic indication. When K is equal to 1, if there are multiple subbands with the same highest channel quality indicator (CQI) value, only the subband corresponding to the first highest CQI in the reported CQI sequence from left to right ordering, or only the subband corresponding to the first highest CQI in the reported CQI sequence from high to low ordering, or only the subband with the largest or smallest subband index corresponding to the highest CQI in the reported CQI sequence is selected. The CQI sequence is included in a CSI report. The transmitting the recovered CSI and / or the recovered differential CSI of the selected part of the subbands based on the selection of the part of the subbands indicated by the subband indicator comprises selecting a precoding matrix of the part of the subbands of the downlink transmission and representing the precoding matrix with a subband indication bitmap. The recovered CSI is quantized and compressed based on a second quantization and compression manner, and the second quantization and compression manner is transmitting precoding coefficients corresponding to a single polarization direction. The transmitting precoding coefficients corresponding to a single polarization direction comprises that a precoding matrix of the single polarization direction carries a polarization direction indicator to indicate an antenna polarization direction corresponding to a precoding matrix of the downlink transmission. The polarization direction indicator is represented by 1 bit, and the polarization direction indicator is 0 to represent an antenna port of a +45° polarization direction, and the polarization direction indicator is 1 to represent an antenna port of a -45° polarization direction. The recovered CSI is quantized and compressed based on a third quantization and compression manner, and the third quantization and compression manner is quantizing amplitude coefficients and / or phase coefficients of the precoding matrix. The amplitude coefficients and / or the phase coefficients are quantized by 3 bits or 4 bits. The quantizing the amplitude coefficients and / or the phase coefficients of the precoding matrix comprises indicating an amplitude quantization manner of the amplitude coefficients by an amplitude coefficient quantization indicator and / or indicating an amplitude quantization manner of the phase coefficients by a phase coefficient quantization indicator. The amplitude coefficient quantization indicator is represented by 1 bit, and the amplitude coefficient quantization indicator is 0 to represent that the amplitude coefficients are quantized by 3 bits, and the amplitude coefficient quantization indicator is 1 to represent that the amplitude coefficients are quantized by 4 bits. The phase coefficient quantization indicator is represented by 1 bit, and the phase coefficient quantization indicator is 0 to represent that the phase coefficients are quantized by 3 bits, and the phase coefficient quantization indicator is 1 to represent that the phase coefficients are quantized by 4 bits.

54. The artificial intelligence model monitoring method of claim 49, further comprising: calculating a quantization error of the recovered CSI before and after the third quantization and compression manner. The recovered CSI comprises first information and second information, the first information is a set of indicators of quantization and compression of the recovered CSI, and the second information is CSI data of quantization and compression of the recovered CSI. The first information is received by static indication, periodic configuration, semi-static indication, or dynamic indication.

57. The artificial intelligence model monitoring method of claim 55, wherein, When the recovery CSI of the partial subband and / or the recovery differential CSI is selected for transmission based on the preset rule, the data composition of the first information from the most significant bit (MSB) to the least significant bit (LSB) respectively represents: The quantization compression mode indicator respectively corresponds to whether the first quantization compression mode, the second quantization compression mode and the third quantization compression mode are adopted for quantization compression of the CSI information, wherein the first quantization compression mode is to select the recovery CSI of the partial subband and / or the recovery differential CSI for transmission, the second quantization compression mode is to select the precoding coefficient corresponding to a single polarization direction for transmission, and the third quantization compression mode is to quantize the amplitude coefficient and / or the phase coefficient of the precoding matrix; The polarization direction selection indicator indicates the antenna polarization direction corresponding to the precoding matrix of the downlink transmission; The amplitude coefficient quantization indicator indicates the quantization mode of the amplitude coefficient of the precoding matrix element; The phase coefficient quantization indicator indicates the quantization mode of the phase coefficient of the precoding matrix element; The quantization compression similarity indicator indicates the quantized value of the cosine similarity after quantization compression; The second information data size indicator is used to indicate the data bit number of the second information.

58. The artificial intelligence model monitoring method of claim 55, wherein, When the recovery CSI of the partial subband and / or the recovery differential CSI is selected for transmission according to the subband indicator, the data composition of the first information from the most significant bit (MSB) to the least significant bit (LSB) respectively represents: The quantization compression mode indicator respectively corresponds to whether the first quantization compression mode, the second quantization compression mode and the third quantization compression mode are adopted for quantization compression of the CSI information, wherein the first quantization compression mode is to select the recovery CSI of the partial subband and / or the recovery differential CSI for transmission, the second quantization compression mode is to select the precoding coefficient corresponding to a single polarization direction for transmission, and the third quantization compression mode is to quantize the amplitude coefficient and / or the phase coefficient of the precoding matrix; The polarization direction selection indicator indicates the antenna polarization direction corresponding to the precoding matrix of the downlink transmission; The amplitude coefficient quantization indicator indicates the quantization mode of the amplitude coefficient of the precoding matrix element; The phase coefficient quantization indicator indicates the quantization mode of the phase coefficient of the precoding matrix element; The quantization compression similarity indicator indicates the quantized value of the cosine similarity after quantization compression; The second information data size indicator is used to indicate the data bit number of the second information; The subband indicator indicates the selected subband in the form of a bitmap. Receiving the first precoding matrix indicator (PMI) sent by the first device; 59.A method for artificial intelligence model monitoring, performed at a second device, the second device being deployed with a channel state information (CSI) decoder applying an artificial intelligence model, the first device being correspondingly deployed with a CSI encoder applying an artificial intelligence model, the method comprising: And Comparing the first PMI with the second PMI of the second device; Wherein, The second PMI is the corresponding PMI of the recovery CSI output by the CSI decoder of the second device, and the first PMI is the corresponding PMI of the original CSI input by the CSI encoder of the first device. ​ 60. The artificial intelligence model monitoring method of claim 59, wherein, The corresponding PMI of the recovered CSI comprises first information and second information, the first information comprises codebook compression related configuration parameters, and the second information is CSI data subjected to codebook based quantization compression of the recovered CSI.

61. The artificial intelligence model monitoring method of claim 60, wherein, The codebook compression related configuration parameters comprise a codebook type, a selected number of spatial domain bases, a selected number of frequency domain bases, and a maximum number of non-zero coefficients of weighting amplitude coefficients of subband components.

62. The artificial intelligence model monitoring method of claim 61, wherein, The corresponding PMI of the recovered CSI corresponds to a parameter combination of the codebook type.

63. The artificial intelligence model monitoring method of claim 61, wherein, Parameters of the corresponding PMI of the recovered CSI are indicated in the first information.

64. The artificial intelligence model monitoring method of claim 60, wherein, The first information is received through static indication, periodic configuration, semi-static indication, or dynamic indication.

65. An artificial intelligence model monitoring method, executed on a second device, the second device being deployed with a channel state information (CSI) decoder applying an artificial intelligence model, a first device being deployed with a CSI encoder applying an artificial intelligence model, the method comprising: determining corresponding second downlink channel quality information according to recovered CSI; receiving first downlink channel quality information sent by the first device, wherein the first downlink channel quality information is determined based on original CSI; comparing model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information; wherein, the recovered CSI is recovered CSI output by a CSI decoder of the second device, and the original CSI is original CSI input by a CSI encoder of the first device.

66. The artificial intelligence model monitoring method of claim 65, wherein, The first downlink channel quality information and the second downlink channel quality information comprise signal to interference plus noise ratio (SINR), reference signal received power (RSRP), reference signal received quality (RSRQ), received signal strength indication (RSSI), or channel quality indicator (CQI).

67. The artificial intelligence model monitoring method of claim 65, wherein, A difference between the first downlink channel quality information and the second downlink channel quality information is greater than a preset threshold.

68. The artificial intelligence model monitoring method of claim 65, wherein, The comparing of the model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information comprises: comparing a similarity of the first downlink channel quality information and the second downlink channel quality information.

69. The artificial intelligence model monitoring method of claim 65, wherein, The comparing of the model monitoring performance indicators based on the first downlink channel quality information and the second downlink channel quality information comprises: comparing a positive correlation of a cosine similarity between the first downlink channel quality information and a cosine similarity between the original CSI and the recovered CSI.

70. The artificial intelligence model monitoring method of claim 65, wherein, A cosine similarity and / or a normalized mean square error between the first downlink channel quality information and the second downlink channel quality information are used as the model monitoring performance indicators.

71. A wireless communication device, comprising: A processor and a memory, the memory being used to store a computer program, and the processor being used to invoke and run the computer program stored in the memory, to execute the method according to any one of 1 to 70.

Images