Communication method and communication device

Abstract

The invention provides a communication method and a communication device. The communication method comprises the following steps: receiving a plurality of first reference signals sent by network equipment; determining a first data set based on the plurality of first reference signals; the first data set is used for at least one of training, reasoning or performance monitoring of the first model; obtaining a second data set; the second data set is a data set used for training a second model, the second model and the first model are the same in structure, and a first similarity index of the second data set and the first data set meets a first condition; and training the first model based on the second data set and / or the second model to obtain a third model. According to the method, the air interface overhead in the model training process can be reduced.

Description

Technical Field

[0001] The present application relates to the field of communications, and in particular to a communication method and a communication device. Background Art

[0002] Channel state information (CSI) processing is a key technology in wireless communications and an important means of achieving more efficient and reliable networks. With the rapid development of technologies such as artificial intelligence (AI) and machine learning in recent years, models have been applied to CSI processing, including but not limited to CSI prediction and CSI compression.

[0003] The CSI processing model (called the CSI model) is trained based on a training dataset. This training dataset is generated based on reference signals (RS) sent by network devices. However, in today's large-scale multiple-input multiple-output (MIMO) scenarios, constructing the training dataset requires a large number of reference signals, resulting in significant air interface overhead. Therefore, it is necessary to study how to reduce the air interface overhead required for CSI model training. Summary of the Invention

[0004] The present application provides a communication method and a communication device that can reduce the air interface overhead required for model training and save network resources.

[0005] In a first aspect, a communication method is provided. This method can be executed, for example, by a terminal device, or by a component configured in the terminal device (such as a circuit, chip, or chip system), or by a logic module or software that implements all or part of the terminal device's functions. This application is not limited to this. The following description uses a terminal device as an example.

[0006] The method includes: receiving multiple first reference signals sent by a network device; determining a first data set based on the multiple first reference signals; using the first data set for at least one of training, reasoning, or performance monitoring of a first model; obtaining a second data set; the second data set is a data set used for training a second model, the second model has the same structure as the first model, and a first similarity index between the second data set and the first data set meets a first condition; training the first model based on the second data set and / or the second model to obtain a third model.

[0007] The communication method provided in the first aspect of the present application obtains a second data set of a second model with the same structure as the first model, and the similarity index of the second data set and the first data set currently used by the first model meets the first condition, and the two are relatively similar. The first data set can reflect the network scenario in which the current terminal device is located, so the network scenario corresponding to the second data set similar to the first data set is similar to the current network scenario. Therefore, training the first model based on the second data set and / or the second model can not only train a model that meets the needs of the current scenario and achieve the training effect of the model, but also reuse the historical training data set and / or use the knowledge of the second model to replace the supervised learning function, which can reduce the demand for new training data, thereby reducing the demand for network devices to send reference signals, thereby reducing air interface overhead and effectively saving network resources.

[0008] In one possible implementation, the first similarity indicator satisfies the first condition, including: the first similarity indicator is the one with the highest similarity indicated by at least one second similarity indicator; each second similarity indicator represents the degree of similarity between a third data set and the first data set, each third data set is a data set used to train a fourth model, and the fourth model is a model in the terminal device with the same structure as the first model.

[0009] In this implementation, the first condition is a condition related to relative extreme values. In this way, it is possible to screen out a second data set that is relatively most similar to the first data set, and to screen out a historical training data set that best represents the current network scenario or channel environment, thereby making the third model obtained based on the second data set and / or the second model training more compatible with the current scenario, thereby improving the model training effect. In addition, in this implementation, the relative extreme value range of the first condition is the second similarity index, and the second similarity index is the similarity index between the third data set corresponding to the fourth model in the terminal device and the first data set. In other words, this implementation screens the second data set locally from the terminal device, thus eliminating the need to interact with the network device, further reducing air interface overhead, and further reducing network resources.

[0010] In one possible implementation, the first similarity index includes a first Euclidean distance, the first Euclidean distance represents the Euclidean distance between the feature information of the second data set and the feature information of the first data set, and the second similarity index includes a second Euclidean distance, each second Euclidean distance represents the Euclidean distance between the feature information of a third data set and the feature information of the first data set; the first similarity index satisfies the first condition including: the first Euclidean distance is the smallest one of at least one second Euclidean distance.

[0011] In this implementation, the first similarity index can be simply and quickly screened through the Euclidean distance, thereby determining the second data set, improving the screening efficiency, and further improving the model training efficiency.

[0012] In a possible implementation, the feature information includes a distribution feature vector.

[0013] In this implementation, the distribution feature vector can effectively characterize the characteristics of the set, thereby facilitating the determination of the similarity between two sets.

[0014] In one possible implementation, the method also includes: receiving multiple second reference signals sent by the network device; determining a fourth data set based on the multiple second reference signals; training the first model based on the second data set and / or the second model to obtain a third model, including: performing self-distillation on the first model based on the fourth data set and the second model to obtain the third model.

[0015] The fourth data set can be understood as a training data set corresponding to the reference signal received in real time.

[0016] In this implementation, first, a self-distillation approach is used to train the first model based on the second model. This eliminates the need to provide a large teacher model and associated data to network devices, further reducing air interface overhead and conserving network resources. Second, when training the first model based on the second model, a real-time dataset (the fourth dataset) is used to reflect the actual channel state. This reduces air interface overhead during training of the first model while also taking into account real-time scenarios, improving model training effectiveness and enhancing the accuracy of the resulting third model.

[0017] In one possible implementation, the fourth data set includes input data and true values; based on the fourth data set and the second model, the first model is self-distilled to obtain a third model, including: determining an output distribution based on the input data and the first model; determining a soft label based on the input data and the second model; aligning the soft label with the output distribution to obtain a soft loss; using the true value as a hard label, aligning the hard label with the output distribution to obtain a hard loss; determining a loss function based on a weighted sum of the soft loss and the hard loss; and updating the first model to convergence based on the loss function to obtain a third model.

[0018] In this implementation, in the process of self-distillation of the first model based on the second model, the rich supervisory information of the second model is extracted through soft labels, and the true value is used as a hard label, which can reflect the current real channel environment and network scenario. Therefore, the loss function obtained by weighted summation after aligning the soft label and the hard label can not only make full use of the useful information of the second model, but also take into account the current real scenario, thereby reducing the air interface overhead while taking into account the model training effect, thereby improving the performance of the third model.

[0019] In one possible implementation, in the loss function, the weight coefficient corresponding to the soft label is positively correlated with the degree of similarity represented by the first similarity indicator.

[0020] In other words, the higher the degree of similarity represented by the first similarity indicator, the larger the weight coefficient corresponding to the soft label, and the lower the degree of similarity represented by the first similarity indicator, the smaller the weight coefficient corresponding to the soft label. The higher the degree of similarity represented by the first similarity indicator, the more similar the second data set is to the first data set, which means that the second model is more consistent with the current scenario, and the second model is more referenceable, or in other words, the second model has a higher utilization value. Therefore, setting a larger soft loss weight coefficient can make the second model's supervisory and guidance role on the first model account for a larger proportion. In this way, the supervisory and guidance role of the second model is fully utilized, the model training efficiency is improved, the model distillation effect is improved, and the performance of the third model is thereby improved.

[0021] In one possible implementation, the output distribution includes a first probability distribution and a second probability distribution; determining the output distribution based on the input data and the first model includes: inputting the input data into the first model to obtain a first original output; using a first temperature parameter to normalize the first original output to obtain a first probability distribution; the first temperature parameter is greater than 1; using a second temperature parameter to normalize the first original output to obtain a second probability distribution; the second temperature parameter is equal to 1; determining a soft label based on the input data and the second model includes: inputting the input data into the second model to obtain a second original output; using the first temperature parameter to normalize the second original output to obtain a third probability distribution; and using the third probability distribution as a soft label.

[0022] In this implementation, the first original output and the second original output are respectively normalized with a temperature parameter greater than 1 to achieve temperature scaling of the first original output and the second original output, thereby softening the probability distribution and facilitating the extraction of generalized knowledge (or "dark knowledge") of the second model, so that the first model can not only learn the results, but also learn the association information between categories, thereby improving the effectiveness of model training and improving the performance of the third model.

[0023] In a possible implementation, the first temperature parameter is negatively correlated with the degree of similarity represented by the first similarity indicator.

[0024] In other words, the higher the degree of similarity represented by the first similarity index, the smaller the first temperature parameter, and the lower the degree of similarity represented by the first similarity index, the larger the first temperature parameter. The lower the degree of similarity represented by the first similarity index, the less similar the second dataset is to the first dataset. Therefore, setting a larger first temperature parameter can achieve stronger temperature scaling (i.e., increasing the temperature), making the resulting soft labels smoother, thereby more effectively exposing the relationships between categories in the second model, allowing the first model to learn more about these relationships between categories, improving model learning efficiency, enhancing the model distillation effect, and ultimately improving the performance of the third model.

[0025] In one possible implementation, the soft labels are aligned with the output distribution to obtain a soft loss, including: determining the soft loss based on the KL divergence between the first probability distribution and the third probability distribution; and the true value is used as the hard label and the hard label is aligned with the output distribution to obtain a hard loss, including: determining the hard loss based on the cross entropy of the second probability distribution and the true value.

[0026] In this implementation, the soft loss is calculated using KL divergence, which is highly sensitive to differences in non-dominant categories in the second model. This effectively transfers the generalized knowledge of the teacher model, improving model training effectiveness. The hard loss is determined using cross-entropy, which directly optimizes the classification objective and has a high penalty for incorrect predictions. This allows for anchoring hard labels, preventing misleading second-model predictions, and accelerating model convergence.

[0027] In a possible implementation manner, the method further includes: sending first level information to the network device, where the first level information indicates a similarity level corresponding to the first similarity indicator.

[0028] In this implementation, the first level information is sent to the network device to facilitate the network device to control the sending rate of the second reference signal.

[0029] In a possible implementation, the first level information is sent to the network device via the PMO field.

[0030] The PMO field is used to indicate the performance monitoring output of the model. In this implementation, a new value is added to the PMO field to transmit the similarity level. This allows the reuse of existing signaling, reduces signaling types, and simplifies information transmission.

[0031] In a possible implementation, the number of second reference signals corresponds to the first level information.

[0032] In this implementation, first-level information is sent to a network device to instruct it to transmit a number of second reference signals corresponding to the first-level information. In other words, controlling the number of second reference signals transmitted based on the similarity between the first and second datasets effectively balances air interface overhead and model training effectiveness, improving the performance of the trained third model.

[0033] In a possible implementation manner, the method further includes: sending information related to the similarity level division to the network device.

[0034] In this implementation, information related to the similarity level division is sent to the network device so that the network device and the terminal device can synchronize the similarity levels indicated by the similarity levels, thereby facilitating the network device to more accurately determine the sending amount of the second reference signal.

[0035] In a possible implementation, information related to similarity level division is transmitted through the reportparameter field.

[0036] The report parameter field is used to transmit measurement information. In this implementation, by adding a new value to the report parameter field, information related to similarity level division can be transmitted, which can reuse existing signaling, reduce signaling types, and simplify information transmission.

[0037] In a possible implementation, the information related to the similarity level division includes one or more of the following: the number of similarity levels, the interval threshold corresponding to the similarity level, or the interval width of the similarity level.

[0038] In one possible implementation, training the first model based on the second data set and / or the second model to obtain the third model includes: if the first similarity index meets the second condition, training the first model based on the second data set and / or the second model to obtain the third model.

[0039] In one possible implementation, the method further includes: if the first similarity index does not meet the second condition, receiving multiple second reference signals sent by the network device, and training the first model based on the multiple second reference signals to obtain a third model.

[0040] In these two implementations, if the second dataset is highly similar to the first dataset, the first model is trained based on the second dataset and / or the second model. Otherwise, the first model is trained using the reference signal sent by the network device in a traditional manner. This ensures effective model training and improves the performance of the trained third model.

[0041] In a possible implementation, the first data set is used for performance monitoring of the first model; and obtaining the second data set includes: if the result of the performance monitoring of the first model does not meet the third condition, obtaining the second data set.

[0042] In this implementation, the training of the first model is triggered in combination with the performance monitoring of the first model, which can realize the automation of model training and optimization, save the power consumption of electronic equipment and improve the model training effect.

[0043] In a second aspect, a communication method is provided. This method can be performed, for example, by a network device, or by components configured within the network device (such as circuits, chips, or chip systems). It can also be implemented by a logic module or software that implements all or part of the network device's functionality. This application is not limited to this. The following description uses a network device (such as a satellite) as an example.

[0044] The method includes: sending a first reference signal to a terminal device; the first reference signal is used to determine a first data set, and the first data set is used for at least one of training, reasoning or performance monitoring of a first model; receiving first level information sent by the terminal device; the first level information indicates a similarity level corresponding to a first similarity indicator, and the first similarity indicator characterizes the degree of similarity between the first data set and the second data set, the second data set is a data set used to train a second model, the second model has the same structure as the first model, the first similarity indicator satisfies a first condition, and the second data set and / or the second model are used to train the first model to obtain a third model.

[0045] In one possible implementation, the first similarity indicator satisfies the first condition, including: the first similarity indicator is the one with the highest similarity indicated by at least one second similarity indicator; each second similarity indicator represents the degree of similarity between a third data set and the first data set, each third data set is a data set used to train a fourth model, and the fourth model is a model in the terminal device with the same structure as the first model.

[0046] In one possible implementation, the first similarity index includes a first Euclidean distance, the first Euclidean distance represents the Euclidean distance between the feature information of the second data set and the feature information of the first data set, and the second similarity index includes a second Euclidean distance, each second Euclidean distance represents the Euclidean distance between the feature information of a third data set and the feature information of the first data set; the first similarity index satisfies the first condition including: the first Euclidean distance is the smallest one of at least one second Euclidean distance.

[0047] In a possible implementation, the feature information includes a distribution feature vector.

[0048] In a possible implementation manner, the method further includes: sending a plurality of second reference signals to the terminal device; the number of the second reference signals corresponds to the first level information.

[0049] In a possible implementation manner, the method further includes: receiving information related to similarity level division sent by the terminal device.

[0050] The second aspect is the implementation on the network device side corresponding to the first aspect. The explanation, supplement and description of the beneficial effects of the first aspect are also applicable to the second aspect and will not be repeated here.

[0051] In a third aspect, a communication device is provided, comprising a processing module and a transceiver module. The transceiver module is configured to receive multiple first reference signals transmitted by a network device; the processing module is configured to: determine a first data set based on the multiple first reference signals; the first data set is used for at least one of training, inference, or performance monitoring of a first model; obtain a second data set; the second data set is used for training a second model, the second model has the same structure as the first model, and a first similarity index between the second data set and the first data set satisfies a first condition; and train the first model based on the second data set and / or the second model to obtain a third model.

[0052] In a fourth aspect, a communication device is provided, comprising a transceiver module. The transceiver module is configured to send a first reference signal to a terminal device; the first reference signal is used to determine a first data set, the first data set being used for at least one of training, inference, or performance monitoring of a first model; and receive first level information sent by the terminal device; the first level information indicating a similarity level corresponding to a first similarity indicator, the first similarity indicator representing a degree of similarity between the first data set and a second data set, the second data set being a data set used to train a second model, the second model having the same structure as the first model, the first similarity indicator satisfying a first condition, and the second data set and / or the second model being used to train the first model to obtain a third model.

[0053] The third and fourth aspects are the device-side implementations corresponding to the first and second aspects. The explanations, supplements and descriptions of the beneficial effects of the first and second aspects are also applicable to the third and fourth aspects and will not be repeated here.

[0054] In a fifth aspect, a communication device is provided, comprising a processor. The processor is coupled to a memory and can be configured to execute instructions or data in the memory to implement the method of any possible implementation of the first aspect. Optionally, the communication device further comprises a memory. Optionally, the communication device further comprises a communication interface, the processor being coupled to the communication interface.

[0055] In one implementation, the communication interface may be a transceiver, or an input / output interface.

[0056] In another implementation, the communication device is a chip configured in a terminal device. When the communication device is a chip configured in a terminal device, the communication interface may be an input / output interface.

[0057] In a sixth aspect, a communication device is provided, comprising a processor. The processor is coupled to a memory and can be configured to execute instructions or data in the memory to implement the method of any possible implementation of the second aspect. Optionally, the communication device further comprises a memory. Optionally, the communication device further comprises a communication interface, the processor being coupled to the communication interface.

[0058] In one implementation, the communication interface may be a transceiver, or an input / output interface.

[0059] In another implementation, the communication device is a chip configured in a satellite. When the communication device is a chip configured in a satellite, the communication interface may be an input / output interface.

[0060] In a seventh aspect, a processor is provided, comprising: an input circuit, an output circuit, and a processing circuit. The processing circuit is configured to receive a signal through the input circuit and transmit a signal through the output circuit, so that the processor executes the method in any possible implementation of any aspect.

[0061] In a specific implementation, the processor may be one or more chips, the input circuit may be an input pin, the output circuit may be an output pin, and the processing circuit may be a transistor, a gate circuit, a trigger, or various logic circuits. The input signal received by the input circuit may be, for example, but not limited to, received and input by a receiver, and the signal output by the output circuit may be, for example, but not limited to, output to and transmitted by a transmitter. The input circuit and the output circuit may be the same circuit, which functions as an input circuit and an output circuit at different times. The embodiments of the present application do not limit the specific implementation of the processor and various circuits.

[0062] In an eighth aspect, a communication device is provided, comprising a processor and a memory. The processor is configured to read instructions stored in the memory and receive signals via a receiver and transmit signals via a transmitter to execute the method of any possible implementation of any of the above aspects.

[0063] Optionally, there are one or more processors and one or more memories.

[0064] In a ninth aspect, a computer program product is provided, which includes: a computer program (also referred to as code, or instructions), which, when executed, enables a computer to execute a method in any possible implementation of any of the above aspects.

[0065] In the tenth aspect, a computer-readable storage medium is provided, which stores a computer program (also referred to as code, or instructions) which, when executed on a computer, enables the computer to execute a method in any possible implementation of any of the above aspects.

[0066] In an eleventh aspect, embodiments of the present application provide a chip system comprising one or more processors configured to retrieve and execute instructions stored in a memory, thereby executing the method of any of the above aspects or any possible implementations of each aspect. The chip system may be composed of a chip or may include a chip and other discrete devices.

[0067] Among them, the chip system may include an input circuit or interface for sending information or data, and an output circuit or interface for receiving information or data.

[0068] In a twelfth aspect, a communication system is provided, comprising the aforementioned terminal device and network device. Optionally, the communication system may further comprise other devices that communicate with the terminal device and / or the network device. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a schematic diagram of the structure of a communication system provided in an embodiment of the present application; Figure 2 This is a schematic diagram of a CSI model design process provided in an embodiment of the present application; Figure 3 This is a flow chart of a communication method provided in an embodiment of the present application; Figure 4 This is a flow chart of another communication method provided in an embodiment of the present application; Figure 5 This is a flow chart of another communication method provided in an embodiment of the present application; Figure 6 This is a flow chart of another communication method provided in an embodiment of the present application; Figure 7 This is a schematic diagram of a communication method provided in an embodiment of the present application; Figure 8 This is a flow chart of another communication method provided in an embodiment of the present application; Figure 9 This is a schematic diagram of the self-distillation principle provided in an embodiment of the present application; Figure 10 This is a schematic diagram of the structure of a communication device provided in an embodiment of the present application; Figure 11 is a structural diagram of another communication device provided in an embodiment of the present application; Figure 12 This is a structural diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings.

[0071] The technical solutions provided in this application can be applied to various communication systems, such as: global system for mobile communications (GSM) system, general packet radio service (GPRS), wireless local area network (WLAN), long term evolution (LTE) system, LTE frequency division duplex (FDD) system, LTE time division duplex (TDD) system, sidelink communication system, universal mobile telecommunication system (UMTS), worldwide interoperability for microwave access (WiMAX) communication system, non-terrestrial network (NTN) communication system, fifth generation (5G) mobile communication system or new radio access technology (NR). Among them, 5G mobile communication system can include non-standalone (NSA) and / or standalone (SA) networking. The technical solutions provided in this application can also be applied to future communication systems. This application is not limited to this.

[0072] Figure 1 Schematic diagram of a communication system 100 used in an embodiment of the present application. The communication system 100 may include network devices, such as Figure 1 The communication system 100 may also include terminal devices, such as Figure 1 The terminal device 120 is shown. The network device 110 and the terminal device 120 can communicate via a wireless link.

[0073] Figure 1 The example shows one network device 110 and one terminal device 120. Optionally, the communication system 100 may also include multiple network devices and / or multiple terminal devices.

[0074] The network equipment in this application may be an access network, core network equipment, or other network-side equipment. Access network equipment is sometimes also referred to as an access node. Access network equipment has wireless transceiver functions and is used to communicate with terminals. Access network equipment includes but is not limited to base stations (base stations) in the above-mentioned communication systems, evolved NodeBs (eNodeBs), transmission reception points (TRPs), next generation NodeBs (gNBs) in 5G mobile communication systems, access network equipment or modules of access network equipment in open access networks (open RAN, ORAN) systems, satellites in NTN communication systems, base stations in future mobile communication systems, or access nodes in WiFi systems, etc. Access network equipment may also be modules or units that can implement some of the functions of a base station. Access network equipment may be a macro base station (such as Figure 1 110a in), micro base stations or indoor stations (such as Figure 1 110b in the figure), a relay node or a donor node, or a wireless controller in a cloud radio access network (CRAN) scenario. Optionally, the access network device may also be a server, a wearable device, or an on-board device. For example, the access network device in the vehicle to everything (V2X) technology may be a road side unit (RSU). The multiple access network devices in the communication system may be base stations of the same type or different types. The base station may communicate with the terminal or communicate with the terminal through a relay station. The terminal may communicate with multiple base stations in different access technologies. The embodiments of the present application do not limit the specific technology and specific device form adopted by the access network device. In the present application, the access network device is referred to as the network device.

[0075] In this application, the device for implementing the function of a network device can be a network device, or a device that can support the network device to implement the function, such as a processor, circuit, chip, or chip system, etc. The device can be installed in the network device or connected to the network device for use. In the technical solution provided in this application, the technical solution provided in this application is described by taking the device for implementing the function of a network device as an example.

[0076] The terminal device in this application may be a wireless terminal device capable of receiving network device scheduling and instruction information. A wireless terminal device may be a device that provides voice and / or data connectivity to a user, a handheld device with wireless connectivity, or other processing device connected to a wireless modem. For example, a terminal device may communicate with one or more core networks or the Internet via a radio access network (RAN). A terminal device may also be referred to as a terminal, user equipment (UE), mobile station, or mobile terminal. The terminal device can be widely used in various scenarios, such as device-to-device (D2D), vehicle-to-everything (V2X) communication, machine-type communication (MTC), Internet of Things (IoT), ultra-reliable low-latency communication (URLLC), virtual reality, augmented reality, industrial control, autonomous driving, telemedicine, smart grid, smart furniture, smart office, smart wearables, smart transportation, smart city, or satellite communication. The terminal may be a mobile phone, tablet computer, computer with wireless transceiver function, wearable device, vehicle, aircraft (such as drone, helicopter, airplane), hot air balloon, ship, robot, robotic arm, or smart home device, etc. The embodiments of the present application do not limit the form of the terminal device.

[0077] In this application, the device for implementing the function of a terminal device can be a terminal device, or a device that can support the terminal device to implement the function, such as a processor, circuit, chip, chip system, etc. The device can be installed in the terminal device or connected to the terminal device for use. In the technical solution provided in this application, the technical solution provided in this application is described by taking the terminal device as an example in which the device for implementing the function of the terminal device is a terminal device.

[0078] The access network equipment and / or the terminal can be fixed or movable. The access network equipment and / or the terminal can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on the water surface; they can also be deployed on airplanes, balloons and artificial satellites in the air. The embodiments of the present application do not limit the application scenarios of the access network equipment and terminals. The access network equipment and the terminal equipment can be deployed in the same scenario or different scenarios. For example, the access network equipment and the terminal equipment are deployed on land at the same time; or, the access network equipment is deployed on land and the terminal equipment is deployed on the water surface, etc., and no further examples are given.

[0079] In practical applications, multiple network devices can collaborate to assist terminals in achieving wireless access, with different network devices each implementing portions of a base station's functionality. For example, a network device can be a centralized unit (CU), a distributed unit (DU), a CU-control plane (CP), a CU-user plane (UP), or a radio unit (RU). The CU and DU can be separate or included in the same network element, such as a baseband unit (BBU). The RU can be included in a radio frequency device or radio unit, such as a remote radio unit (RRU), an active antenna unit (AAU), or a remote radio head (RRH).

[0080] In different systems, the CU (or CU-CP and CU-UP), DU, or RU may have different names, but those skilled in the art will understand their meanings. For example, in the ORAN system, the CU may also be called an O-CU (Open CU), the DU may also be called an O-DU, the CU-CP may also be called an O-CU-CP, the CU-UP may also be called an O-CU-UP, and the RU may also be called an O-RU. Any of the CU (or CU-CP, CU-UP), DU, and RU in this application may be implemented as a software module, a hardware module, or a combination of software and hardware modules. The CU (or CU-CP and CU-UP), DU, and RU can implement different protocol layer functions.

[0081] To facilitate understanding of the embodiments of the present application, a brief description of the terms and related technologies involved in the present application is first given. Optionally, the interpretation of some terms can also refer to the interpretation in the 3rd Generation Partnership Project (3GPP) standard protocol.

[0082] 1. Model In an embodiment of the present application, the model may include an AI model or a machine learning (ML) model. Among them, the AI model is an algorithm or computer program that can implement AI functions. The model characterizes the mapping relationship between the input and output of the model. In an embodiment of the present application, the AI model can be a neural network, a linear regression model, a decision tree model, a support vector machine (SVM), a Bayesian network or a Q learning model.

[0083] 2. Channel state information (CSI) In communication systems (such as LTE or NR), network equipment can determine and schedule the resources, modulation and coding scheme (MCS), and precoding configuration of the downlink data channel of a terminal device based on CSI. CSI is a type of channel information that reflects channel characteristics and quality.

[0084] CSI may include at least one of the following: rank indicator (RI), channel quality indicator (CQI), precoding matrix indicator (PMI), CSI-RS resource indicator (CRI), layer indicator (LI), reference signal receiving power (RSRP), or signal to interference plus noise ratio (SINR). The signal to interference plus noise ratio may also be called the signal to interference plus noise ratio.

[0085] CSI can be obtained through measurement. CSI measurement involves the receiver determining channel information based on a reference signal (RS) sent by the transmitter. This involves estimating the channel information using a channel estimation method. Reference signals include, for example, the channel state information reference signal (CSI-RS) or the synchronization signal block (SSB).

[0086] CSI can also be predicted. With the development of AI and ML technologies, models can be used to predict CSI in the communications field. For example, AI or ML models can be used to predict CSI at future times based on CSI measured at historical times.

[0087] In addition, the model can also be used for other CSI processing, such as CSI compression or decompression.

[0088] In the embodiment of the present application, the model used to process CSI is referred to as a CSI model.

[0089] 3. Use Cases of the CSI Model Network devices and / or terminal devices can perform CSI-related processing through the model. The following examples provide several CSI model use cases from the perspectives of beam management (BM) and CSI feedback enhancement.

[0090] 1) Beam management The network device and / or terminal device may predict CSI for beam management based on the CSI model. The CSI used for beam management may include, for example, the signal quality of a reference signal, such as Layer 1 reference signal received power (L1-RSRP). In embodiments of the present application, the CSI model predicted for beam management may be referred to as the CSI model for beam management.

[0091] Two CSI model use cases for beam management (hereinafter referred to as beam management use cases, also called BM-cases) are given as examples.

[0092] a. Beam management use case 1: Beam prediction in the spatial domain based on the CSI model.

[0093] The application of beam management use case 1 to a terminal device is used as an example for illustration, and the CSI model used by beam management use case 1 is referred to as CSI model 1.

[0094] Specifically, in beam management use case 1, the terminal device can perform spatial domain downlink beam prediction for reference signal set A based on CSI model 1 according to the measurement results of reference signal set B. Reference signal set A is larger than reference signal set B, that is, reference signal set A includes more reference signals than reference signal set B.

[0095] Optionally, the network device may send a reference signal set B to the terminal device. The terminal device measures the layer 1 signal quality (e.g., L1-RSRP) of each reference signal in reference signal set B. The terminal device then inputs the L1-RSRP of each reference signal in reference signal set B into CSI model 1. CSI model 1 outputs the L1-RSRP of each reference signal in reference signal set A. Alternatively, CSI model 1 may output the probability that each reference signal in set A has the largest L1-RSRP value in set A. Reference signal set A is larger than reference signal set B. Therefore, CSI model 1 is used to predict the signal quality of some reference signals (reference signals included in set A but not in set B), thereby enabling prediction of some beams. That is, beam prediction in the spatial domain is achieved.

[0096] b. Beam management use case 2: Time domain beam prediction based on the CSI model.

[0097] Continuing with the application of beam management use case 2 to a terminal device as an example, the CSI model used in beam management use case 2 is referred to as CSI model 2.

[0098] Specifically, in beam management use case 2, the terminal device can perform time-domain downlink beam prediction for reference signal set C based on CSI model 2, based on the measurement results of reference signal set D. Reference signal set D includes reference signals for M transmission occasions in the past (i.e., historical time period). Reference signal set C includes reference signals for N future time periods.

[0099] Optionally, the network device can send a reference signal to the terminal device. At a certain time T, the terminal device uses the reference signals of M transmission opportunities before time T (i.e., the historical time period) as a reference signal set D and measures the layer 1 signal quality (e.g., L1-RSRP) of each reference signal in reference signal set D. The terminal device then inputs the L1-RSRP of each reference signal in reference signal set D into CSI model 2. CSI model 2 outputs the L1-RSRP of each reference signal in reference signal set C. Alternatively, CSI model 2 can output the probability that each reference signal in set C has the highest L1-RSRP value in set C. Reference signal set C includes reference signals for N times after time T (i.e., in the future). CSI model 2 is used to predict the signal quality of reference signals at future times, thereby enabling beam prediction for future times, that is, beam prediction in the time domain.

[0100] It should be noted that the L1-RSRP of each reference signal in the reference signal set D may be used as the entire input of the CSI model 2 or as part of the input, and the specific number may be determined according to actual needs.

[0101] 2) CSI feedback enhancement Terminal devices and / or network devices can enhance CSI feedback based on the CSI model, such as predicting CSI feedback information or compressing or decompressing the CSI to be fed back. CSI feedback information includes, but is not limited to, the channel matrix, channel singular value vectors, and precoding matrix indicator (PMI).

[0102] Two use cases of a CSI model for CSI feedback enhancement (hereinafter referred to as CSI feedback enhancement use cases) are exemplified.

[0103] a. CSI feedback enhancement use case 1: CSI feedback information prediction based on the CSI model.

[0104] CSI feedback enhancement use case 1 is somewhat similar to beam management use case 2, and also involves time domain prediction.

[0105] The following description continues with the application of CSI feedback enhancement use case 1 to a terminal device as an example, and the CSI model used in CSI feedback enhancement use case 1 is referred to as CSI model 3.

[0106] Specifically, in CSI feedback enhancement use case 1, the terminal device can predict the CSI feedback information at the next N moments based on CSI model 3 according to the CSI measurement results of the reference signal at M moments in the historical time period (i.e., the past).

[0107] Optionally, the network device can send a reference signal to the terminal device. The terminal device measures the reference signal at M time instants to obtain CSI measurement results corresponding to the M time instants (e.g., channel matrix, channel singular value vector, or PMI). The terminal device then inputs the CSI measurement results corresponding to the M time instants into CSI model 3, which outputs predicted CSI feedback information (e.g., channel matrix, channel singular value vector, PMI, etc.) for the next N time instants.

[0108] It should be noted that the CSI measurement results of the reference signals at M moments may be used as all or part of the input of the CSI model 3, and the specific number may be determined according to actual needs.

[0109] b. CSI feedback enhancement use case 2: CSI compression based on CSI model.

[0110] The following description continues with the application of CSI feedback enhancement use case 2 to a terminal device as an example, and the CSI model used in CSI feedback enhancement use case 2 is referred to as CSI model 4.

[0111] Specifically, in CSI feedback enhancement use case 2, the terminal device can input the CSI to be fed back into CSI model 4 for compression, and CSI model 4 outputs the CSI compression result. The terminal device can carry the CSI compression result in a CSI report and report it to the network device. The network device can include a CSI reconstruction CSI model. The network device decompresses the CSI compression result using the reconstructed CSI model to obtain the CSI fed back by the terminal device.

[0112] It should be noted that the CSI to be fed back compressed by CSI model 4 can be a measured CSI measurement result, or a CSI prediction result predicted based on the CSI model or other methods, without limitation. Furthermore, the CSI to be fed back can serve as all or part of the input to CSI model 4, depending on actual needs.

[0113] It should be noted that the above AI use cases are only examples and do not represent the only AI use cases for communication systems. Furthermore, a terminal device may support one or more of the above use cases, depending on the terminal device's capabilities.

[0114] 4. Design of CSI Model The design of the CSI model mainly includes data collection, model training, model reasoning and model monitoring.

[0115] For example, Figure 2 A schematic diagram of the CSI model design process is shown in FIG. Figure 2 As shown, the data collection phase is used to collect data required for model training, model inference, or model monitoring. Taking the data collection required for model training as an example, the network device can send a reference signal for model training to the terminal device. The terminal device generates training data for model training based on the reference signal and collects the data to form a training dataset.

[0116] During model training, the terminal device analyzes or trains the initial model based on the training dataset to obtain a CSI model. The CSI model represents the mapping relationship between the model's inputs and outputs. Learning the CSI model through model training is equivalent to learning the mapping relationship between the model's inputs and outputs using the training data.

[0117] During model inference, network devices can also send reference signals to terminal devices for inference. The terminal devices then generate inference data for model inference based on the reference signals. It should be noted that during model inference, the inference data may or may not form an inference dataset, depending on the actual model requirements.

[0118] After generating the inference data, the terminal device uses the CSI model trained through the model training phase to perform inference based on the inference data to obtain an inference result. Specifically, the inference data can be input into the CSI model, and the output can be obtained through the CSI model, and the output is the inference result. Taking the above-mentioned CSI model use case for beam management as an example, the model inference phase also predicts the CSI used for beam management in the spatial domain or time domain based on the CSI model. Taking the above-mentioned CSI feedback enhancement use case 1 as an example, the model inference phase also predicts the CSI feedback information based on the CSI model. Taking the above-mentioned CSI feedback enhancement use case 2 as an example, the model inference phase also performs CSI compression based on the CSI model.

[0119] During the model monitoring phase, network devices can send reference signals for performance monitoring to terminal devices. The terminal devices generate monitoring data based on the reference signals and collect them into a monitoring dataset. The terminal devices and / or network devices monitor the performance of the CSI model based on the monitoring dataset.

[0120] Regarding model performance monitoring, there are three types: In Type 1, the terminal device inputs a monitoring dataset into the CSI model, which then performs inference and outputs a result (called a test result). The terminal device then performs relevant calculations and judgments based on the test result to obtain a performance monitoring output. For example, the performance monitoring output indicates whether the model meets preset requirements. If the model meets the requirements, the performance monitoring output is 1; if not, the performance monitoring output is 0. The terminal device sends the performance monitoring output to the network device.

[0121] In Type 2, the terminal device inputs the monitoring data set into the CSI model, which then outputs the test results. The terminal device then sends the test results to the network device. The network device then performs relevant calculations and judgments based on the test results to generate performance monitoring outputs.

[0122] In Type 3, the terminal device inputs a monitoring dataset into the CSI model, which then outputs test results. The terminal device then performs calculations based on the test results to determine the value of a metric (called a performance indicator) that characterizes the CSI model's performance. The terminal device then sends the performance indicator value to the network device. The network device then performs relevant judgments based on the performance indicator value to generate performance monitoring output.

[0123] It should be noted that in some other embodiments, the model monitoring process can also be performed during the model inference process. Specifically, the terminal device and the network device can monitor the performance of the CSI model based on the inference data or inference results.

[0124] The performance monitoring output of the model can be used to guide model management. For example, when the performance monitoring output of the CSI model does not meet the preset requirements, indicating that the model is unusable, the training dataset can be collected again, and the model training phase can be re-entered to retrain a new CSI model.

[0125] It can be understood that the network device is required to send reference signals in the training, inference and monitoring stages of the CSI model. However, in different stages, the timing, period, or method of configuring the reference signal of the network device may be different. In the embodiments of the present application, the reference signals in different stages can be distinguished.

[0126] In addition, for CSI models with different functions, the reference signals transmitted by the network device can be different, so as to be derived from different reference signals. For example, for the CSI model in the aforementioned CSI feedback enhancement use case, the reference signal transmitted by the network device can include a channel state information reference signal (CSI-RS). For the CSI model used in the aforementioned beam management use case, the reference signal transmitted by the network device can include a CSI-RS and / or SSB.

[0127] It is understandable that in order to improve the accuracy of the model, the training of the CSI model requires a large amount of training data. Figure 2 As can be seen from the corresponding description, training data must be generated based on reference signals sent by network devices. In today's massive MIMO scenarios, training CSI models requires even greater amounts of reference signal data, resulting in significant air interface overhead for constructing training datasets.

[0128] In view of this, the present application provides a communication method, in which a terminal device receives multiple first reference signals transmitted by a network device and determines a first data set based on the multiple first reference signals. The first data set is used for at least one of training, inference, or monitoring of a first model. Subsequently, a second data set is obtained based on the first data set. The second data set is used to train a second model, and the second model has the same structure as the first model. In other words, the second model is a historical model corresponding to the first model, and the second data set is a historical training data set for the second model. At the same time, the similarity index between the second data set and the first data set (referred to as the first similarity index) satisfies a first condition, that is, the second data set is relatively similar to the first data set. The first data set can reflect the current channel state and the network scenario currently in which the terminal device is located. The second data set is relatively similar to the first data set, indicating that the network scenario corresponding to the second data set is similar to the current network scenario, and further indicates that the second model trained based on the second data set is relatively close to the model required for the current network scenario. Therefore, training the first model based on the second data set and / or the second model can not only achieve the purpose of model training, but also reuse the second data set or utilize the experience and instructions of the second model to reduce the demand for new training data, thereby reducing the demand for network devices to send reference signals, thereby reducing air interface overhead and effectively saving network resources.

[0129] The following describes the solution provided by this application in detail in conjunction with the corresponding flowcharts. It will be understood that the schematic flowcharts provided in this application primarily use different devices (e.g., terminal devices, network devices) as examples of the execution entities of the interaction diagrams to illustrate the method, but this application does not limit the execution entities of the interaction diagrams. For example, the device (e.g., terminal device, network device) in the schematic flowcharts may also be a chip, chip system, or processor that supports the device to implement the method, or may be a logic module or software that can implement all or part of the functions of the device.

[0130] For a unified explanation here, in the interaction process of the embodiment of the present application, the message or signaling interaction involved can adopt the message or signaling in the standard, or it can be a newly introduced message or signaling, and the embodiment of the present application does not make specific limitations on this.

[0131] Figure 3 It is a schematic diagram of a communication method 300 according to an embodiment of the present application. It can be understood that Figure 3 The terminal device in can be Figure 1 Any terminal device in the network can also refer to a device in the terminal device (such as a processor, chip, or chip system, etc.). A network device can be Figure 1 Any access network device in the access network device may also refer to a device in the access network device (such as a processor, chip, or chip system, etc.). Figure 3 As shown, the method 300 includes the following steps: S310: The network device sends a plurality of first reference signals to the terminal device. Correspondingly, the terminal device receives the plurality of first reference signals sent by the network device.

[0132] S320, the terminal device determines a first data set based on multiple first reference signals; the first data set is used for at least one of training, reasoning or performance monitoring of the first model.

[0133] The first dataset is used to perform at least one of training, inference, or performance monitoring on the model. That is, the first dataset can be a training dataset for the first model, an inference dataset for the first model, a monitoring dataset for the first model, or a multi-purpose dataset for the first model. For example, the first dataset can be both an inference dataset for the first model and a performance monitoring dataset for the first model. In short, the first dataset is a dataset related to the first model and can serve as input in the training, inference, or performance monitoring phases of the first model.

[0134] Optionally, the first model may be an AI model or an ML model. The function or type of the first model is not limited. For example, the first model may be the CSI model for beam management described above, or a CSI model for CSI feedback enhancement. The first model may be a unilateral model on the terminal device side, or the terminal device side portion of a terminal device-network device bilateral model.

[0135] It is understood that the data in the first data set may be different depending on the function of the first model. For example, if the first model is the aforementioned CSI model for beam management, the first data set may include layer 1 signal qualities (e.g., L1-RSRP) corresponding to multiple reference signals; if the first model is the CSI model in feedback enhancement use case 1, the first data set may include CSI measurement results corresponding to multiple time instants (e.g., channel matrix, channel singular value vector, or PMI); if the first model is the CSI model in feedback enhancement use case 2, the first data set may include CSI to be fed back, which may be a measured CSI measurement result or a CSI prediction result predicted based on the CSI model or other means.

[0136] S330, the terminal device obtains a second data set; the second data set is a data set used to train the second model, the second model has the same structure as the first model, and the first similarity index between the second data set and the first data set meets the first condition.

[0137] In the embodiments of the present application, the same structure of two models can be understood as the same initial model used to train the two models. Specifically, the same structure of two models can be understood as one or more of the following parameters of the two models being the same: network architecture, input and output, or parameter scale. The same network architecture can include the same number of network layers, the type of each layer, and the connection method. For example, both models include N-layer networks, each network layer is a CNN network, and the connection method is residual connection. The same input and output can include the same type of input data and the same type of output data. For example, the input of both models is a channel matrix and the output is a PMI. The same parameter scale can include the same number of neurons or channels in each layer of the network and the same dimension of the weight matrix. It should be understood that two models with the same structure have the same function. However, the same structure of two models does not mean that the model parameters of the two models are the same, nor does it mean that the performance of the two models is the same. Two models with the same structure may have differences in training data, parameter initialization, hyperparameters, regularization, etc., and the model parameters and performance of the two models may differ.

[0138] In this embodiment of the present application, the second dataset is used to train the second model. In other words, the second model is trained based on the second dataset, or in other words, the second dataset is the training dataset for the second model. The first model and the second model have the same structure, and thus the first model and the second model can be understood as copies (or shared parameters) of the same initial model at different training stages. Essentially, these two models can be considered the results of the same initial model in different training states.

[0139] For ease of understanding, in some embodiments of the present application, the first model may also be referred to as the current model or the model to be trained, indicating that the first model is the model currently to be trained (or optimized). Models trained in historical time periods based on the same initial model as the first model are collectively referred to as historical models. The second model belongs to the historical model. Accordingly, the first dataset can be referred to as the current dataset, the training datasets of the historical models can be collectively referred to as the historical training dataset, and the second dataset belongs to the historical training dataset.

[0140] It's understandable that different network scenarios and channel environments in which terminal devices operate result in different reference signals received from network devices, and thus different datasets are collected. For example, when a terminal device is in an indoor scenario versus an outdoor scenario, the channel environment differs, resulting in different datasets. Similarly, when a terminal device is in line-of-sight (LoS) versus non-line-of-sight (NLoS) scenarios, the channel environment also differs, resulting in different datasets. Furthermore, when a terminal device moves at different speeds (for example, in high-speed rail versus non-high-speed rail scenarios), the channel environment also differs, resulting in different datasets. Conversely, the datasets collected by a terminal device can reflect the channel environment and the network scenario in which the terminal device operates. Furthermore, models trained based on different datasets can be applied to different channel environments or network scenarios.

[0141] In an embodiment of the present application, the similarity index (referred to as the first similarity index) between the second data set and the first data set satisfies the first condition. The first condition is a condition that characterizes the degree of similarity between the two data sets. The fact that the first similarity satisfies the first condition indicates that the second data set has a high degree of similarity to the first data set, that is, the second data set is relatively similar to the first data set. Therefore, it means that the channel environment corresponding to the second data set is relatively similar to the channel environment corresponding to the first data set, and that the network scenario corresponding to the second data set is relatively similar to the network scenario corresponding to the first data set. Furthermore, the model trained based on the second data set is similar to the model trained based on the first data set, and can be applied to the same or similar channel environments or network scenarios. Therefore, since the second model is trained based on the second data set, the second model is more compatible with the current network scenario.

[0142] A similarity metric is a quantitative measure of the degree of similarity between two datasets. Optionally, the similarity metric can be similarity or a distance metric. Similarity is positively correlated with the degree of similarity; that is, the higher the similarity, the more similar the two datasets are. Optionally, similarity can be represented by one or more of cosine similarity, intersection-over-union (intersection-over-union), Pearson correlation coefficient, normalized mean square error (NMSE), or square of generalized cosine similarity (SGCS).

[0143] Distance is negatively correlated with similarity; that is, the smaller the distance, the more similar the two datasets are. Alternatively, the distance can be related to a distance parameter such as Euclidean distance, Manhattan distance, or edit distance. It is understood that distance and similarity can be converted to each other; for example, similarity can be the inverse of Euclidean distance.

[0144] Optionally, the first condition may be a threshold-related condition. For example, the first condition may include: the similarity index is greater than a first index threshold; or the first condition may include: the similarity index is less than the first index threshold. The first index threshold may be predefined or configured or indicated by the network device.

[0145] Optionally, the first condition may also be a condition related to a relative extreme value. For example, the first condition may include: the similarity index is the largest one among the candidate similarity indexes; or the first condition may include: the similarity index is the smallest one among the candidate similarity indexes.

[0146] It is understood that the specific content of the first condition may vary depending on the similarity indicator used to represent the degree of similarity. For example, if the similarity indicator is similarity, the first condition may include: the similarity is greater than a first similarity threshold; or the first condition may include: the similarity indicator is the largest among the candidate similarities. If the similarity indicator is distance, the first condition may include: the distance is less than a first distance threshold; or the first condition may include: the similarity indicator is the smallest among the candidate distances.

[0147] S340: The terminal device trains the first model based on the second data set and / or the second model to obtain a third model.

[0148] In one embodiment, the first model may be trained based on the second data set to obtain a third model.

[0149] Optionally, the terminal device may use the second dataset as the entire training dataset to train the first model, thereby obtaining a third model. For example, the second dataset may be input into the first model to obtain a third model. Alternatively, the second dataset may be used as part of the training dataset to train the first model, thereby obtaining a third model. For example, the network device may further transmit a second reference signal to the terminal device, and the terminal device may generate training data (which may be referred to as new training data) based on the second reference signal. The dataset formed by the second dataset and the new training data is then input into the first model, thereby obtaining a third model. For another example, the network device may also use the dataset formed by the second dataset and the first dataset as the training dataset, thereby obtaining a third model.

[0150] As analyzed above, the second dataset is relatively similar to the first, indicating that the network scenario corresponding to the second dataset is close to the current actual network scenario. Therefore, a third model that closely matches the current scenario can be trained based on the second dataset, achieving the purpose of training the first model. Furthermore, reusing the second dataset to train the first model reduces the need for new training data, thereby reducing the need for network devices to send reference signals, lowering air interface overhead, and conserving network resources.

[0151] In another embodiment, the terminal device may also train the first model based on the second model to obtain a third model. Optionally, the terminal device may use the second model as a teacher model and the first model as a student model to train the third model through knowledge distillation. This process can be understood as the second model supervising and guiding the first model, so that the first model learns the "experience" and "knowledge" of the second model, or in other words, transfers the "experience" and "knowledge" of the second model to the first model.

[0152] As analyzed above, the second dataset is relatively similar to the first, indicating that the network scenario corresponding to the second dataset is close to the current actual network scenario. Therefore, the second model trained based on the second dataset is a good match for the current network scenario. Therefore, the "experience" and "knowledge" of the second model are transferred to the first model, making the first model applicable to the current network scenario, thus achieving the purpose of training the first model. Furthermore, in this process, learning from the "experience" and "knowledge" of the second model can replace at least part of the supervised learning function, reducing the demand for new training data. This, in turn, reduces the need for network devices to send reference signals, lowering air interface overhead and conserving network resources.

[0153] In another embodiment, the terminal device can also train the first model based on the second dataset and the second model to obtain a third model. Optionally, during knowledge distillation based on the second model, the second dataset can be used as part or all of the training dataset to train the first model. This reduces the demand for new training data by replacing supervised learning functions and reusing training data, further reducing the need for network devices to send reference signals, further reducing air interface overhead, and further conserving network resources.

[0154] In summary, in the embodiment of the present application, a historical training data set (i.e., a second data set) for a second model with the same structure as the first model is obtained, and the similarity index between the second data set and the first data set currently used by the first model satisfies the first condition, and the two are relatively similar. The first data set can reflect the network scenario in which the current terminal device is located, and thus the network scenario corresponding to the second data set similar to the first data set is similar to the current network scenario. Therefore, training the first model based on the second data set and / or the second model can not only train a model that meets the requirements of the current scenario and achieve the model training effect, but also reuse the historical training data set and / or use the knowledge of the second model to replace the supervised learning function, which can reduce the demand for new training data, thereby reducing the demand for network devices to send reference signals, thereby reducing air interface overhead and effectively saving network resources.

[0155] The following is an illustrative description of the application scenarios of the method provided in this application.

[0156] In one application scenario, the communication method provided in an embodiment of the present application can be used for model retraining. That is, the first model is a model that has been trained at least once. In this case, when the first model needs to be retrained, the above steps S310 to S340 can be started to retrain the first model to obtain a third model.

[0157] Of course, in another application scenario, this method can also be used for the initial training of the model, and is compatible with traditional model training methods. That is, the first model can be an initial model that has not been trained. In this application scenario, the above steps S310 to S330 can be started when the first model needs to be trained. Since this is the first training of the first model, the second data set cannot be obtained. Optionally, when the second data set cannot be obtained, the terminal device can train the first model according to the traditional model training method, that is, the network device sends a second reference signal for training the first model to the terminal device, the terminal device generates a training data set based on the received second reference signal, and trains the first model based on the training data set.

[0158] In the following embodiments of this application, the retraining of the model is mainly taken as an example to illustrate the first data sets for different purposes and the corresponding scenarios.

[0159] In one embodiment, the first data set is used to train the first model, or in other words, the first data set is used to train the first model. For example, when the first model needs to be retrained, referring to the above steps S310 and S320, the network device can send a certain number of first reference signals for training the first model to the terminal device, and the terminal device receives the first reference signal and generates a first data set for training the first model based on the first reference signal. The first data set can be used as part or all of the training data set of the first model, and referring to the above step S340, the first model is trained based on the first data set, the second data set and / or the second model. In one embodiment, the training data set also includes the second data set obtained by the above step S330. That is, the terminal device can use the first data set and the second data set as training data sets to retrain the first model and obtain a third model.

[0160] In one embodiment, the first data set is used to infer the first model, or in other words, the first data set is used for inference of the first model. For example, referring to the above steps S310 and S320, the network device may send a first reference signal for inference of the first model to the terminal device, the terminal device receives the first reference signal, and generates a first data set for inference of the first model based on the first reference signal. The terminal device inputs the first data set into the first model for model inference. The terminal device may determine the performance of the first model based on the model inference result. When the performance of the first model is poor and needs to be retrained, the terminal device may obtain the second data set according to the above step S330. Thereafter, referring to the above step S340, the first model is retrained based on the second data set and / or the second model to obtain a third model.

[0161] In one embodiment, the first data set is used to monitor the performance of the first model, or in other words, the first data set is used to monitor the performance of the first model. Figure 4 In this scenario, the communication method provided by the embodiment of the present application may include: S310: The network device sends a first reference signal for first model performance monitoring to the terminal device. Correspondingly, the terminal device receives the first reference signal for first model performance monitoring sent by the network device.

[0162] S320: The terminal device generates a first data set for first model performance monitoring based on the first reference signal.

[0163] Afterwards, the terminal device performs performance monitoring on the first model based on the first data set. Specifically, this may include: S410: The terminal device inputs a first data set into a first model, and the first model outputs a test result.

[0164] S420: The terminal device calculates the performance index of the first model based on the test result.

[0165] S430: The terminal device determines whether the performance index of the first model satisfies a third condition. If the performance index of the first model satisfies the third condition, step S440 is executed; if the performance index of the first model does not satisfy the third condition, step S330 is executed.

[0166] The third condition can be the aforementioned preset requirement. The third condition is used to characterize the quality of the model's performance. If the model's performance indicators meet the third condition, the model's performance is good, indicating that the model is usable, and a performance monitoring output of 1 can be output. If the model's performance indicators do not meet the third condition, the model's performance is poor, indicating that the model is unusable and requires retraining, and a performance monitoring output of 0 can be output.

[0167] Optionally, the third condition may be a threshold-type condition. For example, the third condition may include: the performance indicator is greater than or equal to the second indicator threshold. The specific content of the third condition may be different depending on the type of performance indicator. In a specific embodiment, taking the first model as an example of a model for predicting PMI, the performance indicator may be the SGCS of the PMI predicted by the first model and the PMI measured. Optionally, an SGCS threshold may be set. The third condition may include: SGCS is greater than the SGCS threshold. If the SGCS of the first model is greater than the SGCS threshold, it indicates that the performance of the first model is better, indicating that the first model is available, and a performance monitoring output of 1 may be output. If the SGCS of the first model is less than or equal to the SGCS threshold, it indicates that the performance of the first model is poor, indicating that the first model is unavailable, and a performance monitoring output of 0 may be output.

[0168] S440: The terminal device sends a performance monitoring output 1 to the network device. Correspondingly, the network device receives the performance monitoring output 1 sent by the terminal device.

[0169] The performance monitoring output 1 indicates that the performance indicator of the first model meets the third condition, or in other words, indicates that the first model is available.

[0170] After step S440 , the process may return to step S310 to perform the next round of performance monitoring on the first model.

[0171] S330: The terminal device obtains a second data set.

[0172] S340: The terminal device trains the first model based on the second data set and / or the second model to obtain a third model.

[0173] If the performance indicator of the first model does not meet the third condition, before or after step S330, the method may further include: S450: The terminal device sends a performance monitoring output of 0 to the network device. Correspondingly, the network device receives the performance monitoring output 0 sent by the terminal device.

[0174] Performance monitoring output 0 indicates that the performance indicator of the first model does not meet the third condition, or in other words, indicates that the first model is unavailable. It should be noted that in this embodiment, performance monitoring output 1 and performance monitoring output 0 are merely examples used to illustrate different performance monitoring outputs and do not limit the representation of the performance monitoring output.

[0175] That is to say, when the first data set is used for performance monitoring of the first model, if the performance monitoring output indicates that the first model is unavailable and the first model needs to be retrained, the terminal device can obtain the second data set according to the above step S330. Thereafter, according to step S340, the first model is retrained based on the second data set and / or the second model to obtain a third model. In other words, the training of the first model is triggered in combination with the performance monitoring of the first model, which can realize the automation of model training and optimization, save the power consumption of electronic equipment and improve the model training effect.

[0176] As can be seen from the above embodiments, the communication method provided by the embodiments of the present application can be applied to the initial training scenario of the model, and can also be applied to the retraining scenario of the model. Moreover, the method can obtain the second data set based on the reference signal of the model training phase, the reference signal of the model reasoning phase, and the reference signal of the model performance monitoring phase, which has a wide range of applications and high flexibility.

[0177] See also Figure 5 In some embodiments, before step S340, in which the terminal device trains the first model based on the second data set and / or the second model to obtain the third model, the method may further include: S350, the network device sends multiple second reference signals to the terminal device; correspondingly, the terminal device receives the multiple second reference signals sent by the network device.

[0178] S360: The terminal device determines a fourth data set based on the multiple second reference signals.

[0179] In step S340, the terminal device trains the first model based on the second data set and / or the second model to obtain the third model, which may include: S341: The terminal device performs self-distillation on the first model based on the fourth data set and the second model to obtain a third model.

[0180] The fourth data set can be understood as a training data set corresponding to the reference signal received in real time.

[0181] In this embodiment, first, a self-distillation approach is used to train the first model based on the second model. This eliminates the need to provide a large teacher model and related data to network devices, further reducing air interface overhead and conserving network resources. Second, when training the first model based on the second model, a real-time dataset (the fourth dataset) is used to reflect the actual channel state. This reduces air interface overhead during training of the first model while also taking into account real-time scenarios, improving model training effectiveness and enhancing the accuracy of the trained third model.

[0182] In some embodiments, the terminal device may set some similarity levels to represent the degree of similarity. The similarity levels may be used to guide the network device to send the number of true values of the second reference signal. Specifically, the following is described: The terminal device may divide the value of the similarity index into intervals, with each interval corresponding to a similarity level. Different similarity levels indicate different degrees of similarity. The method of dividing the similarity levels, the correspondence between similarity levels and similarity levels, etc., can be combined with the specific type of similarity index and set according to actual needs, and this application does not impose specific limitations.

[0183] As a possible implementation, after obtaining the second data set, the terminal device may further determine the similarity level corresponding to the first similarity indicator. For ease of description, the similarity level corresponding to the first similarity indicator is referred to as first level information. The terminal device may transmit the first level information to the network device. The network device may transmit a corresponding number of second reference signals to the terminal device based on the degree of similarity indicated by the first level information. In other words, the number of second reference signals transmitted corresponds to the first level information.

[0184] In some embodiments, the terminal device may also send information related to the similarity level division to the network device so that the network device and the terminal device can synchronize the degree of similarity indicated by the similarity level, thereby facilitating the network device to more accurately determine the amount of the second reference signal sent.

[0185] Optionally, the information related to the similarity level division may include one or more of the following: the number of similarity levels, the interval threshold corresponding to the similarity level, or the interval width of the similarity level.

[0186] In a specific implementation, the higher the degree of similarity indicated by the similarity level, the fewer second reference signals the network device sends to the terminal device; the lower the degree of similarity indicated by the similarity level, the more second reference signals the network device sends to the terminal device. The higher the degree of similarity, the closer the network scenario or channel state corresponding to the second data set is to the current network scenario or channel state, and the stronger the referenceability of the second data set and the second model. Therefore, fewer second reference signals can be sent, ensuring the model training effect while reducing the air interface overhead as much as possible and saving network resources. The lower the degree of similarity, the less close the network scenario or channel state corresponding to the second data set is to the current network scenario or channel state, and the weaker the referenceability of the second data set and the second model. Therefore, more second reference signals can be sent to improve the model training effect.

[0187] In this embodiment, the sending amount of the second reference signal is controlled based on the similarity between the first data set and the second data set, which can effectively take into account and balance the air interface overhead and the model training effect, and improve the performance of the trained third model.

[0188] In addition, in some embodiments, after obtaining the second data set, the terminal device may select whether to train the first model based on the second data set based on the first similarity indicator. For example, if the first similarity indicator satisfies the second condition, the first model is trained based on the second data set and / or the second model as shown in step S340 above. If the first similarity indicator does not satisfy the second condition, the first model is trained based on the reference signal transmitted by the network device in a conventional manner.

[0189] Optionally, the second condition may characterize that the degree of similarity indicated by the similarity index is high. Optionally, the second condition may be a threshold-related condition. For example, the second condition may include: the similarity index is greater than the second similarity index threshold, or the similarity index is less than the second similarity index threshold. Specifically, the second condition may be different depending on the type of similarity index. For example, when the similarity index is similarity, the second condition may include: the first similarity is greater than the second similarity threshold; when the similarity index is distance (such as Euclidean distance), the second condition may include: the first Euclidean distance is less than the second distance threshold.

[0190] That is, if the second dataset is highly similar to the first dataset, the first model is trained based on the second dataset and / or the second model. Otherwise, the first model is trained using the reference signal sent by the network device in a traditional manner. This ensures effective model training and improves the performance of the trained third model.

[0191] Below, we take the first data set used for performance monitoring of the first model as an example, and describe the above implementation methods with reference to the accompanying drawings. Figure 4 and Figure 6 In one implementation, after step S330, the method may further include: S610: The terminal device determines first level information, where the first level information represents a similarity level corresponding to a first similarity indicator.

[0192] In the above S450, the terminal device sends the performance monitoring output 0 to the network device, including: S451: The terminal device sends the performance monitoring output 0, the first level information, and information related to similarity level division to the network device. Correspondingly, the network device receives the performance monitoring output 0 and the first level information sent by the terminal device.

[0193] Afterwards, the method further comprises: S350: The network device sends multiple second reference signals to the terminal device. The amount of the second reference signals sent corresponds to the first level information. In response, the terminal device receives the multiple second reference signals sent by the network device.

[0194] S360: The terminal device determines a fourth data set based on the multiple second reference signals.

[0195] S620, the terminal device determines whether the first similarity index meets the second condition; if the first similarity index meets the second condition, execute step S341; if the first similarity index does not meet the second condition, execute step S550.

[0196] S341: The terminal device performs self-distillation on the first model based on the fourth data set and the second model to obtain a third model.

[0197] S630: The terminal device trains the first model according to the fourth data set to obtain a third model.

[0198] It can be understood that step S341 and step S630 are steps performed under different first similarity indicators, and therefore the corresponding first-level information is different, and the number of second reference signals sent by the network device is different, so the number of data in the fourth data set in step S341 and step S630 is different.

[0199] Regarding the transmission of similarity level information from a terminal device to a network device, including information related to the similarity level classification and the similarity level, the network device may configure or instruct the network device on the transmission method for this information, or predefine the transmission method for this information. The transmission method includes the signaling used, fields, etc. When the terminal device needs to transmit similarity level classification information, it transmits it to the network device based on the transmission method.

[0200] In a specific embodiment, the CSI report framework can be reused, the CSI report configuration (CSI-ReportConfig) can be modified, and new values can be added to the preset fields in the CSI report configuration to implement the transmission of similarity level information. This can reuse existing signaling, reduce signaling types, and simplify information transmission.

[0201] For example, the CSI reporting configuration may be modified as follows: CSI-ReportConfig ::= { reportConfigId = 1, … reportparameter = M- / / Report information related to similarity level classification reportQuantity = PMO- / / Report performance monitoring output and similarity level … } The PMO field is a field used to indicate the performance monitoring output (performance monitoring output) of the model. Indicates the similarity level, specifically the first level information mentioned above. The similarity level is added to the PMO field to realize the transmission of the similarity level and the reuse of the PMO field.

[0202] In the above report parameter field, M and Represents the relevant information of the similarity level division, for example, M can be the number of similarity levels, The report parameter field is a field used to transmit measurement information. In this embodiment, similarity level classification related information is reported through the report parameter field, thereby realizing the reuse of the report parameter field.

[0203] See also Figure 7 In summary, the communication method provided in this embodiment may include two parts. The first part is: selecting a similar second data set and a second model based on the first data set, such as steps S310 to S330 and related content. The second part is: training the first model based on the second data set and / or the second model, such as step S340 and related content.

[0204] For ease of understanding, the first and second parts are further explained below.

[0205] 1. Part 1.

[0206] First, the first similarity index will be described.

[0207] In one embodiment, the first similarity indicator satisfies the first condition, including: the first similarity indicator being the one indicating the highest degree of similarity among at least one second similarity indicator. Each second similarity indicator represents the degree of similarity between a third dataset and the first dataset, each third dataset is a dataset used to train a fourth model, and the fourth model is a model in the terminal device having the same structure as the first model.

[0208] In other words, the models stored in the terminal device that have the same structure as the first model are collectively referred to as the fourth model. The fourth model can be understood as a historical model stored in the terminal device. The training dataset used to train the fourth model is referred to as the third dataset. The third dataset can be understood as a historical training dataset stored in the terminal device.

[0209] In other words, the second dataset is a collection of historical training datasets local to the terminal device. Specifically, the second dataset is selected based on the similarity index between each historical training dataset and the first dataset. The similarity indexes between the historical training datasets and the first dataset are collectively referred to as the second similarity index. The first similarity index is the one of the second similarity indexes that indicates the highest degree of similarity. In other words, the second dataset is the historical training dataset (i.e., the third dataset) in the terminal device that is most similar to the first dataset.

[0210] In this implementation, the first condition is a condition related to relative extreme values. This allows the selection of a second dataset that is relatively most similar to the first dataset, and a historical training dataset that best represents the current network scenario or channel environment. This allows the third model trained based on the second dataset and / or the second model to better match the current scenario, thereby improving the model training effect. Furthermore, in this implementation, the relative extreme value range of the first condition is the historical training dataset in the terminal device, meaning that the second dataset is selected locally on the terminal device without the need for interaction with the network device, further reducing air interface overhead and network resources.

[0211] In one embodiment, the similarity indicator includes an SGCS. Specifically, the first similarity indicator includes a first SGCS, where the first SGCS represents the SGCS between the feature information of the second dataset and the feature information of the first dataset. The second similarity indicator includes second SGCSs, where each second SGCS represents the SGCS between the feature information of a third dataset and the feature information of the first dataset. The first similarity indicator satisfies the first condition including: the first SGCS is the largest of the at least one second SGCS.

[0212] Optionally, the feature information of the dataset may be a feature vector of the dataset. Optionally, the feature vector of the dataset may be calculated based on singular value decomposition (SVD).

[0213] In another embodiment, the similarity index includes a Euclidean distance. Specifically, the first similarity index includes a first Euclidean distance, where the first Euclidean distance represents the Euclidean distance between feature information of the second data set and feature information of the first data set. The second similarity index includes second Euclidean distances, where each second Euclidean distance represents the Euclidean distance between feature information of a third data set and feature information of the first data set. The first similarity index satisfies the first condition including: the first Euclidean distance is the smallest of at least one second Euclidean distance.

[0214] In this embodiment, the first similarity index can be simply and quickly screened through the Euclidean distance, thereby determining the second data set, improving the screening efficiency, and further improving the model training efficiency.

[0215] Optionally, the characteristic information of the data set may be a characteristic distribution vector of the data set. The distribution characteristic vector can effectively characterize the characteristics of the set, thereby facilitating the determination of the similarity between two sets.

[0216] Optionally, the feature distribution vector of the data set can be calculated based on kernel principal component analysis (KPCA).

[0217] Next, taking the first similarity index including the first Euclidean distance as an example, the process of obtaining the second data set is further explained.

[0218] Assume that the network device sends N first reference signals to the terminal device, where N is a positive integer. The first reference signal is, for example, a CSI-RS. The terminal device determines a first data set based on the N first reference signals. The first data set includes, for example, N original channel matrices H corresponding to the N first reference signals. The first data set is represented as ,but .in , Indicates the number of receiving antennas of the terminal device, Indicates the number of transmit antennas of a network device.

[0219] Assume that the terminal device includes P fourth models, where P is a positive integer. The P fourth models correspond to P third data sets respectively. The pth third data set is expressed as , .

[0220] 1) The terminal device calculates the first data set based on the KPCA method The distribution characteristic vector of .

[0221] First dataset The distribution eigenvector of Specifically, the calculation process can be as follows: a. Define the radial basis function, as shown in formula (1): (1).

[0222] in, is the preset parameter, Represents the pairwise channel matrix in the data set The local features between .

[0223] For the first data set In terms of For the first data set The pairwise original signal matrix The local features between .

[0224] b. Composition distribution characteristic matrix , as shown in formula (2): (2).

[0225] c. Distribution characteristic matrix Perform eigenvalue decomposition and take the front Maximum eigenvalue Constructing the first data set The distribution characteristic vector of .

[0226] 2) The terminal device calculates the distribution feature vectors of each third data set based on the KPCA method.

[0227] The pth third data set The distribution characteristic vector of .

[0228] Optionally, N samples can be sampled from each third data set, and the distribution feature vector of the third data set can be calculated based on the N samples. This not only simplifies the calculation, but also facilitates comparison with the distribution feature vector of the first data set, thereby improving the accuracy of the similarity index calculation.

[0229] The process of calculating the distribution feature vector of the third data set based on N samples is the same as that of calculating the distribution feature vector of the first data set. The distribution characteristic vector of The process is the same and will not be repeated here.

[0230] The distribution feature vectors of P third data sets can be expressed as .

[0231] 3) The terminal device calculates the distribution feature vectors of each third data set and the first data set respectively The distribution characteristic vector of The Euclidean distance between them is P, and the second Euclidean distance is obtained.

[0232] Take the pth third data set The distribution characteristic vector of For example, it can be calculated according to formula (3) and Euclidean distance between : (3) The smaller the value, the more representative the first dataset. With the The more similar the characteristics of the third dataset are.

[0233] The P second Euclidean distances can be expressed as: .

[0234] 4) The smallest of the P second Euclidean distances is determined as the first Euclidean distance, and the third data set corresponding to the first Euclidean distance is determined as the second data set.

[0235] The first Euclidean distance is expressed as ,but .

[0236] is the smallest of the P second Euclidean distances, that is, The corresponding third data set is most similar to the first data set, so The corresponding third data set is determined as the second data set.

[0237] In this embodiment, the distribution eigenvectors of the datasets are calculated based on the KPCA method, and the similarity between the two datasets is determined by calculating the Euclidean distance between the two distribution eigenvectors. This method can simply, quickly, and effectively determine the one of the P third datasets that is closest to the first dataset as the second dataset, thereby training the first model based on the second dataset and / or the second model, improving data acquisition efficiency and, in turn, model training efficiency.

[0238] Then, taking the first similarity index including the first Euclidean distance as an example, the division of similarity levels and the determination of first level information are explained.

[0239] In one embodiment, M-1 interval critical values may be preset, which are expressed as: , M is an integer greater than 1. The relationship between the critical values of M-1 intervals is: The critical values of these M-1 intervals are all greater than 0. Among them, is the maximum interval critical value. The M-1 interval critical values are used to divide the M Euclidean distance intervals, which are: ), [ ),…,[ ),…,[ ).in, , m is an integer. Optionally, among the M-1 interval critical values, the difference between the critical values of two adjacent intervals is equal. That is, among the M Euclidean distance intervals, the interval widths of the first M-1 Euclidean distance intervals can be equal. For example, the interval widths of the first M-1 Euclidean distance intervals can all be equal to .

[0240] Different Euclidean distance intervals correspond to different similarities. Optionally, each Euclidean distance interval can be quantified into a quantized value, which represents the similarity of the Euclidean distance interval. The quantized value can be called a similarity level. The similarity level can achieve quantification of the similarity level by stage, simplify similarity management, and facilitate identification and transmission. In a specific embodiment, it can be used Each bit of information is used to quantify each Euclidean distance interval and obtain the corresponding similarity level. ), the corresponding similarity level is expressed as , .

[0241] Determining the similarity level corresponding to a certain Euclidean distance means determining the similarity level corresponding to the Euclidean distance interval to which the Euclidean distance belongs. For example, assuming the first Euclidean distance The Euclidean distance interval is [ ), then the first Euclidean distance The corresponding similarity level, that is, the first level information is .

[0242] It can be understood that the similarity level based on Euclidean distance is that the lower the level, the higher the similarity, and the higher the level, the lower the similarity. , the smaller m is, the higher the similarity is, and the larger m is, the lower the similarity is. The similarity level can be used to adjust the transmission amount of the second reference signal used to train the first model. Specifically, after the terminal device determines the first level information, it can send the first level information to the network device. The network device adjusts the configuration of the second reference signal resource according to the first level and controls the transmission amount of the second reference signal. In this way, based on the similarity between the second data set and the first data set, an appropriate number of second reference signals are sent, which can not only reduce the air interface overhead and save network resources, but also ensure the training effect of the first model training, and achieve a balance between the air interface overhead and the training effect.

[0243] Next, the second condition is further explained by taking the case where the first similarity index includes the first Euclidean distance as an example.

[0244] In one embodiment, when the similarity index includes the Euclidean distance, the first similarity index meeting the second condition may include: the first Euclidean distance Less than the second distance threshold. The second distance threshold can be expressed as That is to say, if , indicating that the terminal device does not have a training data set similar to the current scenario locally, and there is no historical model that closely matches the current scenario, then the first model is trained based on the reference signal sent by the network device in a traditional way; if , indicating that there is a training dataset (second dataset) similar to the current scenario locally on the terminal device, and there is a historical model (second model) that is more closely matched with the current scenario. The second dataset is used as the training dataset, and the second model is used as the self-distillation reference model to train the first model through the self-distillation method.

[0245] In one embodiment, the second distance threshold Can be less than or equal to the maximum interval critical value .

[0246] 2. Part 2.

[0247] Here, taking the self-distillation of the first model based on the fourth data set and the second model as an example, the training process of the first model is furthered.

[0248] Optionally, the fourth dataset may include input data and ground truth (hereinafter referred to as ground truth). The input data is used as input to the model, and the ground truth represents the actual value or standard value of the real-time channel state and is used as a label.

[0249] See also Figure 8 In one embodiment, in step S341, the terminal device performs self-distillation on the first model based on the fourth data set and the second model to obtain a third model, including: S810, the terminal device determines an output distribution based on the input data and the first model; S820, the terminal device determines a soft tag based on the input data and the second model; S830: The terminal device aligns the soft labels with the output distribution to obtain a soft loss. At step S840, the terminal device uses the true value as a hard label and aligns the hard label with the output distribution to obtain a hard loss. S850, the terminal device determines a loss function according to a weighted sum of the soft loss and the hard loss; S860: The terminal device updates the model parameters of the first model based on the loss function until the first model converges to obtain a third model.

[0250] Optionally, when performing self-distillation on the first model, the final output of the first model may be distilled, or the process output within the first model may be distilled.

[0251] In this embodiment, in the process of self-distillation of the first model based on the second model, rich supervisory information of the second model is extracted through soft labels, and the true value is used as a hard label, which can reflect the current real channel environment and network scenario. Therefore, the loss function obtained by weighted summation after aligning the soft label and the hard label can not only make full use of the useful information of the second model, but also take into account the current real scenario, thereby reducing the air interface overhead while taking into account the model training effect, thereby improving the performance of the third model obtained.

[0252] In one embodiment, the output distribution includes a first probability distribution and a second probability distribution. Step S810, determining the output distribution based on the input data and the first model, includes: a. Inputting input data into the first model to obtain a first original output; b. normalizing the first original output using a first temperature parameter to obtain a first probability distribution, where the first temperature parameter is greater than 1; c. Using a second temperature parameter, normalize the first original output to obtain a second probability distribution, where the second temperature parameter is equal to 1.

[0253] In one embodiment, the above step S820, determining the soft label based on the input data and the second model, includes: a. Input the input data into the second model to obtain a second original output; b. normalizing the first original output using the first temperature parameter to obtain a third probability distribution; c. Use the third probability distribution as a soft label.

[0254] In these two embodiments, the first original output and the second original output are respectively normalized with a temperature parameter greater than 1 to achieve temperature scaling of the first original output and the second original output, thereby softening the probability distribution and facilitating the extraction of generalized knowledge (or "dark knowledge") of the second model, so that the first model can not only learn the results, but also learn the association information between categories, thereby improving the effectiveness of model training and improving the performance of the third model.

[0255] In one embodiment, the first temperature parameter is negatively correlated with the degree of similarity represented by the first similarity indicator.

[0256] In other words, the higher the degree of similarity represented by the first similarity index, the smaller the first temperature parameter, and the lower the degree of similarity represented by the first similarity index, the larger the first temperature parameter. The lower the degree of similarity represented by the first similarity index, the less similar the second dataset is to the first dataset. Therefore, setting a larger first temperature parameter can achieve stronger temperature scaling (i.e., increasing the temperature), making the resulting soft labels smoother, thereby more effectively exposing the relationships between categories in the second model, allowing the first model to learn more about these relationships between categories, improving model learning efficiency, enhancing the model distillation effect, and ultimately improving the performance of the third model.

[0257] In one embodiment, the step S830 of aligning the soft labels with the output distribution to obtain the soft loss includes: aligning the soft labels with the first probability distribution to obtain the soft loss.

[0258] Alternatively, the soft loss can be determined using the KL divergence between the first probability distribution and the soft labels (i.e., the third probability distribution). KL divergence is highly sensitive to differences in the non-dominant classes in the second model and can effectively transfer the generalization knowledge of the teacher model. Therefore, calculating the soft loss using KL divergence can improve model training performance.

[0259] In one embodiment, the above step S840, using the true value as the hard label and aligning the hard label with the output distribution to obtain the hard loss, includes: using the true value as the hard label and aligning the hard label with the second probability distribution to obtain the hard loss.

[0260] Alternatively, a hard loss can be calculated by calculating the cross entropy between the hard label (i.e., the true value) and the second probability distribution. Determining the hard loss using cross entropy directly optimizes the classification objective, imposes a high penalty on incorrect predictions, and thus anchors the hard label, preventing misleading input from the second model and accelerating model convergence.

[0261] In one embodiment, the weight coefficient corresponding to the soft tag is positively correlated with the degree of similarity represented by the first similarity indicator.

[0262] In other words, the higher the degree of similarity represented by the first similarity indicator, the larger the weight coefficient corresponding to the soft label, and the lower the degree of similarity represented by the first similarity indicator, the smaller the weight coefficient corresponding to the soft label. The higher the degree of similarity represented by the first similarity indicator, the more similar the second data set is to the first data set, which means that the second model is more consistent with the current scenario, and the second model is more referenceable, or in other words, the second model has a higher utilization value. Therefore, setting a larger soft loss weight coefficient can make the second model's supervisory and guidance role on the first model account for a larger proportion. In this way, the supervisory and guidance role of the second model is fully utilized, the model training efficiency is improved, the model distillation effect is improved, and the performance of the third model is thereby improved.

[0263] For ease of understanding, the first model and the second model are taken as an example to illustrate the models used to predict PMI. That is, the outputs of the first model and the second model are the predicted PMI. In 3GPP Release 18, the PMI can be generated based on the enhanced Type II codebook for predicted PMI and the corresponding codebook parameters. Among them, the enhanced Type II codebook for predicting PMI is also called the Type II Doppler-r18 codebook, etc. The PMI contains the configuration of multiple discrete parameters, such as beam indicators, wideband amplitude indicators, etc., and can therefore be regarded as a classification task. The goal of the classification task can be understood as selecting the optimal parameter value combination of the PMI at the prediction moment from the codebook configuration.

[0264] Optionally, define the raw outputs (output logits) of the first and second models as . Original output represents the unnormalized score of the first model or the second model for each PMI option in the codebook. Represents the total number of output categories, that is, the total number of possible output values. For the model predicting PMI, That is, the total number of discrete parameter combinations of each PMI in Type II Doppler-r18.

[0265] For the sake of distinction, the original output of the first model is called the first original output, which is expressed as , the original output of the second model is called the second original output, expressed as The function corresponding to the first model is expressed as The function corresponding to the second model is expressed as The model parameters of the first model are expressed as , the model parameters of the second model are expressed as The data set formed by the input data in the fourth data set is expressed as In this embodiment, the input data may be a CSI matrix. The true value in the fourth data set may be expressed as True value Indicates the PMI category, for example, the correct PMI category is set to 1, and the rest of the PMI categories are set to 0. True value It can also be expressed in the Type II Doppler-r18 format.

[0266] See Figure 9 As shown in the schematic diagram of the self-distillation principle, the terminal device performs self-distillation on the first model based on the fourth data set and the second model to obtain a third model, including the following steps: 1) Input the input data into the first model to obtain the first original output .

[0267] Specifically, the process can be expressed as the following formula (4): (4).

[0268] 2) Input the input data into the second model to obtain the second original output .

[0269] Specifically, the process can be expressed as the following formula (5): (5).

[0270] 3) Use the first temperature parameter t to adjust the first original output Normalization is performed to obtain a first probability distribution, where the first temperature parameter t is greater than 1.

[0271] In other words, for the first original output Normalization processing is performed for temperature parameters greater than 1.

[0272] Normalization with a temperature parameter greater than 1 can also be called temperature scaling or softening.

[0273] In one embodiment, the value range of the first temperature parameter t can be .

[0274] 4) Using the second temperature parameter, the first original output Normalization is performed to obtain a second probability distribution, where the second temperature parameter is 1.

[0275] In other words, for the second original output Perform normalization with a temperature parameter of 1. Normalization with a temperature parameter of 1 is also called standard normalization, and the result is a standard probability distribution.

[0276] 5) Use the first temperature parameter t to adjust the second original output Normalization is performed to obtain a third probability distribution.

[0277] In other words, for the second original output Normalization processing is performed for temperature parameters greater than 1.

[0278] Optionally, you can pass The function is normalized, and the specific expression is as follows: (6) in, represents the cth of the C original outputs, and T represents the temperature parameter. Optionally, in the above steps 3) and 5), the first original output and the second original output The temperature parameter T used for normalization can be the same, which is the first temperature coefficient t, that is, T=t, t is a value greater than 1, to achieve temperature scaling and soften the distribution. In order to facilitate distinction, the first probability distribution is expressed as , the third probability distribution is expressed as .

[0279] In the above step 4), the second temperature coefficient used for normalizing the first original output is 1, that is, T=1, to achieve normalization. For the convenience of description, the second probability distribution is expressed as .

[0280] The above steps 1) to 5) can also be summarized as: forward propagation to generate probability distribution, steps 1) and 2) can be summarized as calculating the original output, and steps 3) and 5) can be probabilistically described as: temperature scaling to soften the distribution.

[0281] 6) Using the third probability distribution as a soft label, determine the soft loss between the first probability distribution and the soft label.

[0282] In other words, the soft labels (i.e., the third probability distribution) are aligned with the first probability distribution to determine the soft loss (or, in other words, the soft loss is aligned with the soft label). The soft loss represents the confidence of the first model's prediction result for the PMI category compared with the soft label.

[0283] In a specific embodiment, the soft labels may be aligned based on the KL divergence to determine the soft loss. For example, the soft loss may be calculated as follows: (7).

[0284] in, Indicates soft loss.

[0285] By aligning the third probability distribution, the rich soft label information output by the second model is used to supervise the training of the first model, so that the first model can learn the knowledge of the second model, or in other words, the knowledge of the second model is transferred to the first model.

[0286] 7) Take the true value as the hard label and determine the hard loss of the second probability distribution and the hard label.

[0287] In other words, the hard labels (i.e., the true values) are aligned with the second probability distribution to determine the hard loss (or, in other words, to align the hard loss with the hard labels). The hard loss represents the confidence of the first model's prediction for the PMI category compared to the hard label.

[0288] In a specific embodiment, the hard loss can be determined by aligning the hard labels based on cross entropy. For example, the calculation of the hard loss can be as follows: (8).

[0289] in, Indicates a hard loss.

[0290] By aligning the true value, the true value is used to supervise the learning of the first model, or in other words, the first model learns the true value and the real network scenario or channel environment information, so that the trained third model matches the real scenario.

[0291] 8) Determine the loss function based on the weighted sum of hard loss and soft loss.

[0292] The loss function is also called the total loss function or the weighted loss function and can be expressed as Alternatively, the loss function can be determined according to the following formula (9): (9).

[0293] in, is the weight coefficient corresponding to the soft loss. is the weight coefficient corresponding to the hard loss.

[0294] In one embodiment, The value range can be In this way, the weight of soft loss is limited to this range to prevent the proportion of soft loss from being too high or too low, which is conducive to balancing the supervisory and guiding role of the second model on the first model and the constraint effect of the current scene's true value data on the first model, improving the model training effect, and improving the matching degree between the obtained third model and the current real scene, thereby improving the accuracy of the third model.

[0295] 9) Based on the loss function, update the model parameters of the first model until the first model converges to obtain the third model.

[0296] Optionally, a gradient descent method can be used to update the model parameters of the first model. Specifically, the model parameters of the first model can be adjusted along the negative gradient direction of the loss function to minimize the value of the loss function. The model parameters that minimize the value of the loss function are also the model parameters when the first model converges, that is, the model parameters of the third model.

[0297] For example, it can be determined according to the following formula (10): (10).

[0298] in, Represents the model parameters when the first model converges, that is, the model parameters of the third model.

[0299] In one embodiment, the weight coefficient corresponding to the soft loss is The first similarity index may be adjusted. Optionally, the first similarity index includes a first Euclidean distance In the case of The first Euclidean distance Negative correlation. That is, the first Euclidean distance The smaller the weight coefficient The larger the first Euclidean distance The larger the weight coefficient The smaller.

[0300] In a specific embodiment, the weight coefficient corresponding to the soft loss is The first Euclidean distance The relationship can be expressed as the following formula (11): (11).

[0301] Among them, k is a constant used to control the weight coefficient With the first Euclidean distance The decay rate of the decay is larger and smaller.

[0302] Formula (11) can be used to quantify and accurately control the weight coefficient corresponding to the soft loss, thereby improving the model training effect.

[0303] In this embodiment, the first Euclidean distance The smaller the value, the more similar the second dataset is to the first dataset, indicating that the second model is more consistent with the current scenario, and the second model is more referenceable, or the second model has higher utilization value. Therefore, a larger soft loss weight coefficient is set. , which can make the supervisory guidance role of the second model on the first model account for a larger proportion. In this way, the supervisory guidance role of the second model can be fully utilized to improve the model training efficiency, improve the model distillation effect, and thus improve the performance of the third model.

[0304] In one embodiment, the first temperature parameter t can be adjusted according to the first similarity index. Optionally, the first similarity index includes a first Euclidean distance In the case of the first temperature parameter t and the first Euclidean distance Positive correlation. That is, the first Euclidean distance The smaller the first temperature parameter t is, the smaller the first Euclidean distance The larger , the larger the first temperature parameter t.

[0305] In a specific embodiment, the first temperature parameter t and the first Euclidean distance The relationship can be expressed as the following formula (12): (12).

[0306] Formula (12) can be used to quantify and accurately control the first temperature parameter, thereby improving the model training effect.

[0307] In this embodiment, the first Euclidean distance The larger the value, the more dissimilar the second data set is to the first data set. Therefore, setting a larger first temperature parameter t can achieve a stronger temperature scaling (i.e., increase the temperature), making the obtained soft labels smoother, thereby being able to expose more relationships between categories in the second model, allowing the first model to learn more about the relationships between categories, improving the model learning efficiency, improving the model distillation effect, and thus improving the performance of the third model.

[0308] It should be understood that Figures 1 to 9 The flowcharts or scenario diagrams shown are only for ease of understanding and are not intended to limit the embodiments of the present application to the examples shown in the diagrams. In fact, those skilled in the art will Figures 1 to 9 The examples in can be equivalently transformed to obtain more implementation methods.

[0309] Combined with the above Figures 1 to 9, describes in detail the communication method provided by the embodiment of the present application. Figures 10 to 12 It should be understood that the communication device of the present invention can execute the various communication methods of the above embodiments of the present invention, that is, the specific working processes of the following various products can refer to the corresponding processes in the above method embodiments. In each of the above embodiments, the terminal device may perform some or all of the steps in each embodiment; the network device may perform some or all of the steps in each embodiment. These steps or operations are merely examples, and the embodiments of the present application may also perform other operations or variations of various operations. In addition, the various steps may be performed in a different order as presented in the embodiments, and it is possible that not all of the operations in the embodiments of the present application need to be performed. Moreover, the size of the sequence number of each step does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0310] Figure 10 : is a schematic block diagram of a communication device provided in an embodiment of the present application. Figure 10 As shown, the communication device 1000 may include a processing module 1010 and a communication module 1020. The processing module 1010 may implement corresponding processing functions. The communication module 1020 may implement corresponding communication functions, which may be internal communication functions of the communication device 1000 or communication functions between the communication device 1000 and other devices. Optionally, the communication module 1020 may also be referred to as a communication interface or a transceiver module.

[0311] Optionally, the communication device 1000 further includes a storage module, which can be used to store instructions and / or data; the processing module 1010 can read the instructions and / or data in the storage module, so that the communication device 1000 implements the aforementioned method embodiment.

[0312] In one possible design, the communication device 1000 may correspond to the terminal device in the above method embodiments, or a component configured in the terminal device (such as a circuit, chip, or chip system). The communication device 1000 can be used to execute the steps or processes executed by the terminal device in any of the above method embodiments.

[0313] For example, the communication module 1020 is used to: receive multiple first reference signals sent by the network device; the processing module 1010 is used to: determine a first data set based on the multiple first reference signals; the first data set is used for at least one of training, reasoning or performance monitoring of the first model; obtain a second data set; the second data set is a data set used to train the second model, the second model has the same structure as the first model, and the first similarity index of the second data set and the first data set meets the first condition; train the first model based on the second data set and / or the second model to obtain a third model.

[0314] In one embodiment, the first similarity indicator satisfies the first condition, including: the first similarity indicator is the one with the highest similarity indicated by at least one second similarity indicator; each second similarity indicator represents the degree of similarity between a third data set and the first data set, each third data set is a data set used to train a fourth model, and the fourth model is a model in the terminal device with the same structure as the first model.

[0315] In one embodiment, the first similarity index includes a first Euclidean distance, the first Euclidean distance represents the Euclidean distance between the feature information of the second data set and the feature information of the first data set, and the second similarity index includes a second Euclidean distance, each second Euclidean distance represents the Euclidean distance between the feature information of a third data set and the feature information of the first data set; the first similarity index satisfies the first condition including: the first Euclidean distance is the smallest one of at least one second Euclidean distance.

[0316] In one embodiment, the feature information includes a distribution feature vector.

[0317] In one embodiment, the communication module 1020 is further used to: receive multiple second reference signals sent by the network device; the processing module 1010 is further used to: determine a fourth data set based on the multiple second reference signals; and perform self-distillation on the first model based on the fourth data set and the second model to obtain a third model.

[0318] In one embodiment, the fourth data set includes input data and true values; the processing module 1010 is specifically used to: determine the output distribution based on the input data and the first model; determine the soft label based on the input data and the second model; align the soft label with the output distribution to obtain a soft loss; use the true value as a hard label and align the hard label with the output distribution to obtain a hard loss; determine the loss function based on the weighted sum of the soft loss and the hard loss; and update the first model to convergence based on the loss function to obtain a third model.

[0319] In one embodiment, in the loss function, the weight coefficient corresponding to the soft label is positively correlated with the degree of similarity represented by the first similarity indicator.

[0320] In one embodiment, the processing module 1010 is specifically used to: input the input data into the first model to obtain a first original output; use a first temperature parameter to normalize the first original output to obtain a first probability distribution; the first temperature parameter is greater than 1; use a second temperature parameter to normalize the first original output to obtain a second probability distribution; the second temperature parameter is equal to 1; input the input data into the second model to obtain a second original output; use the first temperature parameter to normalize the second original output to obtain a third probability distribution; and use the third probability distribution as a soft label.

[0321] In one embodiment, the first temperature parameter is negatively correlated with the degree of similarity represented by the first similarity indicator.

[0322] In one embodiment, the processing module 1010 is specifically configured to: determine the soft loss according to the KL divergence between the first probability distribution and the third probability distribution; and determine the hard loss according to the cross entropy between the second probability distribution and the true value.

[0323] In one embodiment, the communication module 1020 is further configured to: send first level information to the network device, where the first level information indicates a similarity level corresponding to the first similarity indicator.

[0324] In one embodiment, the number of the second reference signals corresponds to the first level information.

[0325] In one embodiment, the communication module 1020 is further configured to send information related to similarity level division to the network device.

[0326] In one embodiment, the information related to the similarity level division includes one or more of the following: the number of similarity levels, the interval threshold corresponding to the similarity level, or the interval width of the similarity level.

[0327] In one embodiment, the processing module 1010 is specifically configured to: if the first similarity index satisfies the second condition, train the first model based on the second data set and / or the second model to obtain a third model.

[0328] In one embodiment, the processing module 1010 is further configured to: if the first similarity index does not satisfy the second condition, receive multiple second reference signals sent by the network device, and train the first model based on the multiple second reference signals to obtain a third model.

[0329] In one embodiment, the first data set is used for performance monitoring of the first model; the processing module 1010 is specifically configured to: if the result of the performance monitoring of the first model does not satisfy the third condition, obtain the second data set.

[0330] The above is only an example, and for detailed steps or processes, please refer to the description of the aforementioned embodiments.

[0331] Figure 11 This is a schematic block diagram of another communication device provided in an embodiment of the present application. Figure 11 As shown, the communication device 1100 can include a processing module 1110, and the processing module 1110 can implement corresponding processing functions.

[0332] Optionally, the communication device 1100 further includes a storage module, which can be used to store instructions and / or data; the processing module 1110 can read the instructions and / or data in the storage module, so that the communication device 1000 implements the aforementioned method embodiment.

[0333] Optionally, the communication device 1100 further includes a communication module 1120. The communication module 1120 can implement corresponding communication functions, which can be internal communication functions of the communication device 1100 or communication functions between the communication device 1100 and other devices. Optionally, the communication module 1120 can also be called a communication interface or a transceiver module.

[0334] In one possible design, the communication device 1100 may correspond to the network device in the above method embodiments, or a component configured in the network device (such as a circuit, chip, or chip system). The communication device 1100 can be used to execute the steps or processes executed by the network device in any of the above method embodiments.

[0335] For example, the communication module 1120 is used to: send a first reference signal to a terminal device; the first reference signal is used to determine a first data set, and the first data set is used for at least one of training, reasoning or performance monitoring of a first model; receive first level information sent by the terminal device; the first level information indicates a similarity level corresponding to a first similarity indicator, and the first similarity indicator characterizes the degree of similarity between the first data set and the second data set, the second data set is a data set used to train the second model, the second model has the same structure as the first model, the first similarity indicator satisfies the first condition, and the second data set and / or the second model are used to train the first model to obtain a third model.

[0336] In one embodiment, the first similarity indicator satisfies the first condition, including: the first similarity indicator is the one with the highest similarity indicated by at least one second similarity indicator; each second similarity indicator represents the degree of similarity between a third data set and the first data set, each third data set is a data set used to train a fourth model, and the fourth model is a model in the terminal device with the same structure as the first model.

[0337] In one embodiment, the first similarity index includes a first Euclidean distance, the first Euclidean distance represents the Euclidean distance between the feature information of the second data set and the feature information of the first data set, and the second similarity index includes a second Euclidean distance, each second Euclidean distance represents the Euclidean distance between the feature information of a third data set and the feature information of the first data set; the first similarity index satisfies the first condition including: the first Euclidean distance is the smallest one of at least one second Euclidean distance.

[0338] In one embodiment, the feature information includes a distribution feature vector.

[0339] In one embodiment, the communication module 1120 is further configured to: send a plurality of second reference signals to the terminal device; the number of the second reference signals corresponds to the first level information.

[0340] In one embodiment, the communication module 1120 is further configured to receive information related to similarity level division sent by a terminal device.

[0341] In one embodiment, the processing module 1110 is configured to determine the sending amount of the second reference signal based on the first level information.

[0342] The above is only an example, and for detailed steps or processes, please refer to the description of the aforementioned embodiments.

[0343] Figure 12 1 is another schematic block diagram of a communication device 1200 provided in an embodiment of the present application. The communication device 1200 may be a chip, chip system, or processor, etc., that implements the above-described method in a terminal device or network device. The communication device 1200 may be used to implement the method described in the above-described method embodiment. For details, please refer to the description of the above-described method embodiment.

[0344] like Figure 12 As shown, the communication device 1200 may include one or more processors 1210, which may also be referred to as a processing unit or processing module, and may implement certain control functions. The processor 1210 may be a general-purpose processor or a dedicated processor, for example, a baseband processor or a central processing unit. The baseband processor may be used to process communication protocols and communication data, while the central processing unit may be used to control the communication device 1200 (e.g., base station, baseband chip, user, user chip), execute software programs, and process software program data.

[0345] In an optional design, the processor 1210 may also store instructions and / or data, and the instructions and / or data can be executed by the processor 1210, so that the communication device 1200 executes the method described in the above method embodiment.

[0346] In another optional design, the communication device 1200 may include a communication interface 1220 for implementing receiving and transmitting functions. For example, the communication interface 1220 may be a transceiver circuit, an interface, an interface circuit, or a transceiver. The transceiver circuit, interface, interface circuit, or transceiver for implementing the receiving and transmitting functions may be separate or integrated. The transceiver circuit, interface, interface circuit, or transceiver may be used for reading and writing code / data, or the transceiver circuit, interface, interface circuit, or transceiver may be used for transmitting or delivering signals.

[0347] Optionally, the communication device 1200 may include one or more memories 1230, which may store instructions. These instructions may be executed on the processor 1210, causing the communication device 1200 to perform the method described in the above method embodiment. Optionally, the memory 1230 may also store data. Optionally, the processor 1210 may also store instructions and / or data. The processor 1210 and memory 1230 may be provided separately or integrated together.

[0348] It should be understood that, in one possible design, each step in the method embodiment provided in the present application can be completed by an integrated logic circuit of the hardware in the processor or by instructions in the form of software. The steps of the method disclosed in conjunction with the embodiments of the present application can be directly embodied as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor. The software module can be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0349] In one implementation, the communication device 1200 may correspond to the terminal device in the above-mentioned method embodiment, and may be used to execute the various steps and / or processes performed by the terminal device in the above-mentioned method embodiment. The processor 1210 may be used to execute instructions stored in the memory 1230, and when the processor 1210 executes the instructions stored in the memory, the processor 1210 is used to execute the various steps and / or processes of the above-mentioned method embodiment corresponding to the terminal device.

[0350] In another implementation, the communication device 1200 may correspond to the network device in the above-mentioned method embodiment, and may be used to execute the various steps and / or processes performed by the network device in the above-mentioned method embodiment. The processor 1210 may be used to execute instructions stored in the memory 1230, and when the processor 1210 executes the instructions stored in the memory, the processor 1210 is used to execute the various steps and / or processes of the above-mentioned method embodiment corresponding to the network device.

[0351] It should be understood that the processing device may be one or more chips. For example, the processing device may be a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on chip (SoC), a central processor unit (CPU), a network processor (NP), a digital signal processor (DSP), a microcontroller unit (MCU), a programmable logic device (PLD), or other integrated chips.

[0352] It is understood that the memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM bus RAM (DR RAM). 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.

[0353] Based on the methods provided in the embodiments of the present application, the present application also provides a chip system, which includes one or more processors configured to retrieve and execute instructions stored in a memory, thereby executing the methods of the embodiments of the present application. The chip system can be composed of a chip or can include a chip and other discrete devices.

[0354] Among them, the chip system may include an input circuit or interface for sending information or data, and an output circuit or interface for receiving information or data.

[0355] According to the method provided in the embodiment of the present application, the present application also provides a communication system, which includes the aforementioned network device and terminal device.

[0356] According to the method provided in the embodiments of the present application, the present application also provides a computer program product, which includes: computer program code, which, when running on a computer, enables the computer to execute the various steps or processes executed by the network device and terminal device in any of the aforementioned method embodiments.

[0357] According to the method provided in the embodiments of the present application, the present application also provides a computer-readable storage medium, which stores program code. When the program code runs on a computer, the computer executes the various steps or processes performed by the network device and terminal device in any of the aforementioned method embodiments.

[0358] The computer-readable storage medium may be the aforementioned volatile memory or non-volatile memory, or may include both volatile memory and non-volatile memory.

[0359] In the embodiments of this application, each term and English abbreviation is provided for convenience of description and shall not constitute any limitation to this application. This application does not exclude the possibility of defining other terms that can achieve the same or similar functions in existing or future agreements.

[0360] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. A computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present application are generated in whole or in part.

[0361] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0362] It should be understood that in the various embodiments of the present application, the size of the serial number of each process does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0363] In short, the above is only a preferred embodiment of the technical solution of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this application should be included in the scope of protection of this application.

Claims

1. A communication method, executed by a terminal device, characterized in that: include: receiving a plurality of first reference signals sent by a network device; determining a first data set based on the plurality of first reference signals; The first dataset is used for at least one of training, inference, or performance monitoring of a first model; Obtaining a second data set; The second data set is a data set used to train a second model, the second model has the same structure as the first model, and a first similarity index between the second data set and the first data set satisfies a first condition; The first model is trained based on the second data set and / or the second model to obtain a third model.

2. The method according to claim 1, characterized in that The first similarity index satisfies a first condition, including: The first similarity indicator is the one with the highest similarity indicated by at least one second similarity indicator; each of the second similarity indicators represents the degree of similarity between a third data set and the first data set, and each of the third data sets is a data set used to train a fourth model, and the fourth model is a model in the terminal device with the same structure as the first model.

3. The method according to claim 2, characterized in that The first similarity indicator includes a first Euclidean distance, the first Euclidean distance representing the Euclidean distance between the feature information of the second data set and the feature information of the first data set; the second similarity indicator includes a second Euclidean distance, each second Euclidean distance representing the Euclidean distance between the feature information of the third data set and the feature information of the first data set; The first similarity indicator satisfies a first condition including: the first Euclidean distance is the smallest of at least one of the second Euclidean distances.

4. The method according to claim 3, characterized in that The feature information includes a distribution feature vector.

5. The method according to claim 1, characterized in that The method further comprises: receiving a plurality of second reference signals sent by a network device; determining a fourth data set based on the plurality of second reference signals; The step of training the first model based on the second data set and / or the second model to obtain a third model includes: Based on the fourth data set and the second model, self-distillation is performed on the first model to obtain the third model.

6. The method according to claim 5, characterized in that The fourth data set includes input data and true values; and the self-distillation of the first model based on the fourth data set and the second model to obtain the third model includes: determining an output distribution based on the input data and the first model; determining a soft label according to the input data and the second model; Aligning the soft labels with the output distribution to obtain a soft loss; Using the true value as a hard label, aligning the hard label with the output distribution to obtain a hard loss; determining a loss function based on a weighted sum of the soft loss and the hard loss; Based on the loss function, the first model is updated until convergence to obtain the third model.

7. The method according to claim 6, characterized in that In the loss function, the weight coefficient corresponding to the soft label is positively correlated with the degree of similarity represented by the first similarity indicator.

8. The method according to claim 6, characterized in that The output distribution includes a first probability distribution and a second probability distribution; and determining the output distribution according to the input data and the first model includes: Inputting input data into the first model to obtain a first original output; Normalizing the first original output using a first temperature parameter to obtain the first probability distribution; the first temperature parameter is greater than 1; Normalizing the first original output using a second temperature parameter to obtain the second probability distribution; the second temperature parameter is equal to 1; The determining of the soft label according to the input data and the second model includes: Inputting the input data into the second model to obtain a second original output; Normalizing the second original output using the first temperature parameter to obtain a third probability distribution; The third probability distribution is used as the soft label.

9. The method according to claim 8, characterized in that The first temperature parameter is negatively correlated with the degree of similarity represented by the first similarity index.

10. The method according to claim 8, characterized in that Aligning the soft labels with the output distribution to obtain a soft loss includes: determining the soft loss according to a KL divergence between the first probability distribution and the third probability distribution; The true value is used as a hard label, and the hard label is aligned with the output distribution to obtain a hard loss, including: The hard loss is determined according to a cross entropy between the second probability distribution and the true value.

11. The method according to claim 5, characterized in that The method further comprises: First level information is sent to the network device, where the first level information indicates a similarity level corresponding to the first similarity indicator.

12. The method according to claim 11, characterized in that The number of the second reference signals corresponds to the first level information.

13. The method according to claim 11, characterized in that The method further comprises: Send similarity level classification related information to network devices.

14. The method according to claim 12, characterized in that The information related to the similarity level division includes one or more of the following: the number of similarity levels, the interval threshold corresponding to the similarity level, or the interval width of the similarity level.

15. The method according to claim 1, wherein The step of training the first model based on the second data set and / or the second model to obtain a third model includes: If the first similarity index satisfies a second condition, the first model is trained based on the second data set and / or the second model to obtain the third model.

16. The method according to claim 15, characterized in that The method further comprises: If the first similarity index does not meet the second condition, multiple second reference signals sent by the network device are received, and the first model is trained based on the multiple second reference signals to obtain the third model.

17. The method according to any one of claims 1 to 16, characterized in that The first data set is used for performance monitoring of the first model; The obtaining of the second data set includes: If the performance monitoring result of the first model does not meet the third condition, the second data set is obtained.

18. A communication method, executed by a network device, characterized in that: include: Sending a first reference signal to a terminal device; The first reference signal is used to determine a first data set, the first data set being used for at least one of training, inference, or performance monitoring of a first model; receiving the first level information sent by the terminal device; The first level information indicates a similarity level corresponding to a first similarity indicator, the first similarity indicator characterizes the degree of similarity between the first data set and the second data set, the second data set is a data set used to train a second model, the second model has the same structure as the first model, the first similarity indicator satisfies a first condition, and the second data set and / or the second model is used to train the first model to obtain a third model.

19. The method according to claim 18, characterized in that The method further comprises: Sending multiple second reference signals to the terminal device; the number of the second reference signals corresponds to the first level information.

20. The method according to claim 18 or 19, characterized in that The method further comprises: Receive information related to the similarity level division sent by the terminal device.

21. A communication device, characterized in that: The apparatus comprises at least one processor coupled to a memory, wherein the memory stores a program or instruction, and the processor executes the program or instruction so that the apparatus is configured to perform the method according to any one of claims 1 to 20.

22. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instructions are executed, the computer is caused to perform the method according to any one of claims 1 to 20.

23. A communication system, characterized in that: Comprising a communication device as claimed in claim 21.

24. A chip system, characterized in that: The chip system includes one or more processors, which are used to call and execute instructions stored in the memory from the memory, so that the method according to any one of claims 1 to 20 is executed.

Images