Communication method and communication apparatus

Abstract

The application provides a communication method and a communication device. The communication method comprises: receiving a plurality of first reference signals transmitted by a network device; 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, inference 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 of the second data set and the first data set satisfies a first condition; training the first model based on the second data set and / or the second model to obtain a third model. The method can reduce the air interface overhead in the model training process.

Description

TECHNICAL FIELD

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

[0002] The processing of channel state information (CSI) is a key technology in wireless communication and an important means to achieve a more efficient and reliable network. In recent years, with the rapid development of technologies such as Artificial Intelligence (AI) and machine learning, models have been applied to the processing of CSI, including but not limited to CSI prediction, CSI compression, etc.

[0003] The processing model for CSI (referred to as CSI model) is trained based on a training data set. The training data set is generated according to reference signals (RS) sent by network devices. However, in today's large-scale multiple-input multiple-output (MIMO) scenarios, the construction of the training data set requires a large number of reference signals, thus resulting in a large amount of air interface overhead. Therefore, it is necessary to study how to reduce the air interface overhead required for the training of the CSI model. SUMMARY

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

[0005] In a first aspect, a communication method is provided, which can be executed by a terminal device, or can be executed by a component (such as a circuit, a chip or a chip system, etc.) configured in the terminal device, and can also be implemented by a logic module or software that can implement all or part of the functions of the terminal device. The present application does not limit this. Hereinafter, the terminal device is taken as an example for description.

[0006] The method comprises: 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 data set 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 for training a second model, the second model has the same structure as the first model, and a first similarity index of the second data set and the first data set satisfies 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 application can obtain a second data set of a second model having 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 satisfies a first condition, and the two are relatively similar. The first data set can reflect the network scene in which the terminal device is currently located, and therefore the network scene corresponding to the second data set similar to the first data set is similar to the current network scene. 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 scene and achieve the training effect of the model, but also reduce the demand for new training data by reusing historical training data sets and / or replacing the supervised learning function with the knowledge of the second model, thereby reducing the demand for reference signals transmitted by the network device and further reducing the air interface overhead, thereby effectively saving network resources.

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

[0009] In this implementation, the first condition is a condition related to a relative extreme value, so that the second data set that is relatively most similar to the first data set can be screened, and the historical training data set that best represents the current network scene or channel environment can be screened, so that the third model trained based on the second data set and / or the second model has a higher matching degree with the current scene, and the model training effect is improved. In addition, in this implementation, the relative extreme value range of the first condition is a second similarity index, and the second similarity index is a similarity index between a third data set corresponding to a fourth model in the terminal device and the first data set. That is, this implementation screens the second data set from the terminal device locally, so that interaction with the network device is not required, further reducing the air interface overhead and further reducing network resources.

[0010] In a possible implementation, the first similarity index includes a 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, and the second similarity index includes a second Euclidean distance, each second Euclidean distance representing the Euclidean distance between the feature information of a third data set and the feature information of the first data set. The first condition is satisfied when 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, so as to determine the second data set, improve the screening efficiency, and further improve the model training efficiency.

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

[0013] In the implementation manner, the distribution feature vector can effectively represent the features of the set, thereby facilitating the determination of the similarity between two sets.

[0014] In a possible implementation manner, the method further comprises: receiving a plurality of second reference signals sent by the network device; determining a fourth data set according to the plurality of second reference signals; and training the first model based on the second data set and / or the second model to obtain a third model, comprising: 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 the implementation manner, first, the first model is trained in a self-distillation manner based on the second model, without a large teacher model and related data of the network device, thereby further reducing the air interface overhead and saving network resources. Second, when the first model is trained based on the second model, the real-time data set (the fourth data set) can reflect the real channel state, so that the air interface overhead in the process of training the first model is reduced, and the real-time real scene is also taken into account, thereby improving the model training effect and the accuracy of the third model obtained by training.

[0017] In a possible implementation manner, the fourth data set comprises input data and a true value; the self-distillation on the first model based on the fourth data set and the second model to obtain the third model comprises: determining an output distribution according to the input data and the first model; determining a soft label according to the input data and the second model; aligning the soft label with the output distribution to obtain a soft loss; taking the true value as a hard label, and 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 the third model.

[0018] In the implementation manner, in the process of self-distilling the first model based on the second model, the soft label is used to extract rich supervision information of the second model, and the true value is taken as the hard label, which can reflect the current real channel environment and network scene, so that the loss function obtained by aligning and weighting 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 scene, thereby reducing the air interface overhead while taking into account the model training effect, and improving the performance of the obtained third model.

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

[0020] That is, the higher the similarity degree represented by the first similarity index, the greater the weight coefficient corresponding to the soft label, and the lower the similarity degree represented by the first similarity index, the smaller the weight coefficient corresponding to the soft label. The higher the similarity degree represented by the first similarity index, the more similar the second data set is to the first data set, indicating that the second model is more consistent with the current scene, and the referenceability of the second model is stronger, or the utilization value of the second model is higher. Therefore, setting a larger soft loss weight coefficient can make the supervision and guidance of the second model to the first model account for a larger proportion, so as to fully utilize the supervision and guidance of the second model, improve the model training efficiency, improve the model distillation effect, and further improve the performance of the third model.

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

[0022] In this implementation, by performing normalization processing on the first original output and the second original output by using a temperature parameter greater than 1 respectively, temperature scaling is realized on the first original output and the second original output, so as to soften the probability distribution, facilitate the extraction of the generalization knowledge (or referred to as “dark knowledge”) of the second model, and enable the first model to learn not only the result but also the correlation information between categories, thereby improving the effectiveness of model training and the performance of the obtained third model.

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

[0024] That is, the higher the similarity degree represented by the first similarity index, the smaller the first temperature parameter, and the lower the similarity degree represented by the first similarity index, the greater the first temperature parameter. The lower the similarity degree represented by the first similarity index, the less similar the second data set is to the first data set, and therefore, setting a larger first temperature parameter can realize temperature scaling (i.e., increasing the temperature) with greater intensity, so as to make the obtained soft label smoother, thereby exposing the relationship between categories in the second model more, enabling the first model to learn the relationship between categories more, improving the model learning efficiency, improving the model distillation effect, and further improving the performance of the third model.

[0025] In a possible implementation manner, the soft loss is obtained by aligning the output distribution with the soft label, and the hard loss is obtained by aligning the output distribution with the hard label, including: determining the hard loss according to the cross entropy of the second probability distribution and the true value.

[0026] In the implementation manner, the soft loss is calculated by the KL divergence, the KL divergence is highly sensitive to the difference of the non-dominant classes in the second model, can efficiently transfer the generalization knowledge of the teacher model, and improve the model training effect. The hard loss is determined by the cross entropy, which can directly optimize the classification target and has high punishment for the wrong prediction, so as to anchor the hard label, prevent the misguidance of the second model, and accelerate the convergence speed of the model.

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

[0028] In the implementation manner, the first grade information is sent to the network device, so as to facilitate the network device to control the sending quantity of the second reference signal.

[0029] In a possible implementation manner, the first grade information is sent to the network device through a PMO field.

[0030] The PMO field is a field used to indicate the performance monitoring output of a model. In the implementation manner, the transmission of the similarity grade is implemented by adding a new value in the PMO field, so that the existing signaling can be reused, the signaling type is reduced, and the information transmission is simplified.

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

[0032] In the implementation manner, the first grade information is sent to the network device to instruct the network device to send the second reference signal in a quantity corresponding to the first grade information. That is, the sending quantity of the second reference signal is controlled based on the similarity degree between the first data set and the second data set, which can effectively balance the air interface overhead and the model training effect, and improve the performance of the third model obtained by training.

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

[0034] In the implementation manner, the information related to the division of the similarity grade is sent to the network device, so as to facilitate the network device to synchronize the similarity degree indicated by the similarity grade, and facilitate the network device to more accurately determine the sending quantity of the second reference signal.

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

[0036] The reportparameter field is a field used for transmitting measurement information. In this implementation, the transmission of the information related to the similarity level division is implemented by adding a new value in the reportparameter field, so that the existing signaling can be reused, the signaling types can be reduced, and the information transmission can be simplified.

[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 critical value corresponding to the similarity level, or the interval width of the similarity level.

[0038] In a possible implementation, the first model is trained based on the second data set and / or the second model to obtain a third model, including: if the first similarity index satisfies a second condition, training the first model based on the second data set and / or the second model to obtain the third model.

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

[0040] In the two implementations, in a case where the similarity between the second data set and the first data set is high, the first model is trained based on the second data set and / or the second model, otherwise, the first model is trained based on the reference signal sent by the network device in a traditional manner. In this way, the model training effect can be guaranteed, and the performance of the third model obtained by training can be improved.

[0041] In a possible implementation, the first data set is used for performance monitoring of the first model; and the second data set is obtained, including: if a result of the performance monitoring of the first model does not satisfy a 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, so that the automation of model training and optimization can be implemented, the power consumption of the electronic device can be saved, and the model training effect can be improved.

[0043] In a second aspect, a communication method is provided. The method can be executed by a network device, or can be executed by a component (such as a circuit, a chip, or a chip system) configured in the network device, and can also be implemented by a logic module or software that can implement all or part of the functions of the network device. The present application does not limit this. Hereinafter, the network device (such as a satellite) is taken as an example for description.

[0044] The method comprises: 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, inference 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 index, and the first similarity index represents a similarity degree between the first data set and 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, the first similarity index 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 a possible implementation, the first similarity index satisfying the first condition comprises: the first similarity index being a one with the highest similarity degree indicated by at least one second similarity index; each second similarity index represents a similarity degree between a third data set and the first data set, and each third data set is a data set used for training a fourth model, and the fourth model is a model in the terminal device that has the same structure as the first model.

[0046] In a possible implementation, the first similarity index comprises a first Euclidean distance, the first Euclidean distance represents a Euclidean distance between feature information of the second data set and feature information of the first data set, the second similarity index comprises a second Euclidean distance, and each second Euclidean distance represents a Euclidean distance between feature information of a third data set and the feature information of the first data set; the first similarity index satisfying the first condition comprises: the first Euclidean distance being a one with the smallest value among at least one second Euclidean distance.

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

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

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

[0050] The second aspect is an implementation of a network device side corresponding to the first aspect, and the explanations, supplements and beneficial effects of the first aspect are also applicable to the second aspect, and will not be described again.

[0051] In a third aspect, a communication apparatus is provided, which comprises a processing module and a transceiver module. The transceiver module is configured to receive a plurality of first reference signals transmitted by a network device; and the processing module is configured to determine a first data set based on the plurality of 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 a data set used for training a second model, the second model has the same structure as the first model, a first similarity index of the first data set and the second 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 apparatus is provided, which comprises a transceiver module. The transceiver module is configured to transmit a first reference signal to a terminal device; the first reference signal is used to determine a first data set, the first data set is used for at least one of training, inference or performance monitoring of a first model; receive first level information transmitted by the terminal device; the first level information indicates a similarity level corresponding to a first similarity index, the first similarity index represents a similarity degree between the first data set and 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, the first similarity index 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.

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

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

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

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

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

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

[0059] In another implementation form, the communication device is a chip configured in a satellite. When the communication device is a chip configured in a satellite, the communication interface can 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 performs the method in any possible implementation form of any of the aspects.

[0061] In a specific implementation process, the processor can be one or more chips, the input circuit can be an input pin, the output circuit can be an output pin, and the processing circuit can be a transistor, a gate circuit, a flip-flop, various logic circuits, etc. The input signal received by the input circuit can be received and input by, for example but not limited to, a receiver, the signal output by the output circuit can be output to and transmitted by, for example but not limited to, a transmitter, and the input circuit and the output circuit can be the same circuit which is used as the input circuit and the output circuit at different times. The embodiments of the present application do not limit the specific implementation form 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 can receive a signal through a receiver and transmit a signal through a transmitter to perform the method in any possible implementation form of any of the aspects.

[0063] Optionally, the processor is one or more, and the memory is one or more.

[0064] In a ninth aspect, a computer program product is provided, comprising a computer program (which can also be referred to as code or instructions), which, when executed by a computer, causes the computer to perform the method in any possible implementation form of any of the aspects.

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

[0066] In an eleventh aspect, an embodiment of the present application provides a chip system, which comprises one or more processors configured to invoke and run instructions stored in a memory, so that the method in any one of the above aspects or possible implementation manners of the above aspects is executed. The chip system can be composed of a chip, or can comprise a chip and other discrete devices.

[0067] The chip system can comprise an input circuit or interface for transmitting information or data, and an output circuit or interface for receiving information or data.

[0068] In a twelfth aspect, a communication system is provided, which comprises the terminal device and the network device described above. Optionally, the communication system can further comprise other devices in communication with the terminal device and / or the network device. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 FIG. 1 is a structural schematic diagram of a communication system provided by an embodiment of the present application;

[0070] Figure 2 FIG. 2 is a schematic diagram of a CSI model design process provided by an embodiment of the present application;

[0071] Figure 3 FIG. 3 is a flowchart of a communication method provided by an embodiment of the present application;

[0072] Figure 4 FIG. 4 is a flowchart of another communication method provided by an embodiment of the present application;

[0073] Figure 5 FIG. 5 is a flowchart of yet another communication method provided by an embodiment of the present application;

[0074] Figure 6 FIG. 6 is a flowchart of yet another communication method provided by an embodiment of the present application;

[0075] Figure 7 FIG. 7 is a schematic diagram of a principle of a communication method provided by an embodiment of the present application;

[0076] Figure 8 FIG. 8 is a flowchart of yet another communication method provided by an embodiment of the present application;

[0077] Figure 9 FIG. 9 is a schematic diagram of a self-distillation principle provided by an embodiment of the present application;

[0078] Figure 10 FIG. 10 is a structural schematic diagram of a communication apparatus provided by an embodiment of the present application;

[0079] Figure 11 FIG. 11 is a structural schematic diagram of another communication apparatus provided by an embodiment of the present application;

[0080] Figure 12 is a structural schematic diagram of another communication apparatus provided by an embodiment of the present application. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the present application will be described below with reference to the drawings.

[0082] The technical solutions provided by the present application can be applied to various communication systems, such as a global system for mobile communications (GSM) system, a general packet radio service (GPRS), a wireless local area network (WLAN), a long term evolution (LTE) system, an LTE frequency division duplex (FDD) system, an LTE time division duplex (TDD), a sidelink communication system, a universal mobile telecommunication system (UMTS), a worldwide interoperability for microwave access (WiMAX) communication system, a non-terrestrial network (NTN) communication system, a 5th generation (5G) mobile communication system or a new radio access technology (NR). The 5G mobile communication system can include a non-standalone (NSA) and / or a standalone (SA). The technical solutions provided by the present application can also be applied to future communication systems. The present application is not limited in this regard.

[0083] Figure 1 is a schematic diagram of a communication system 100 to which an embodiment of the present application is applied. The communication system 100 can include a network device, such as a network device 110 shown in Figure 1 . The communication system 100 can also include a terminal device, such as a terminal device 120 shown in Figure 1 . The network device 110 and the terminal device 120 can communicate through a wireless link.

[0084] Figure 1A network device 110 and a terminal device 120 are exemplarily shown. Optionally, the communication system 100 can further include multiple network devices and / or multiple terminal devices.

[0085] The network device in the present application can be a device of a network side such as an access network, a core network device, etc. The access network device is also sometimes referred to as an access node. The access network device has a wireless transceiving function and is used to communicate with a terminal. The access network device includes, but is not limited to, a base station in the above-mentioned communication system, an evolved NodeB (eNodeB), a transmission reception point (TRP), a next generation NodeB (gNB) in a 5G mobile communication system, an access network device or a module of an access network device in an open RAN (ORAN) system, a satellite in an NTN communication system, a base station in a future mobile communication system, or an access node in a WiFi system, etc. The access network device can also be a module or unit capable of realizing part of the function of a base station. The access network device can be a macro base station (such as 110a in FIG. 1), a micro base station or an indoor station (such as 110b in FIG. 1), a relay node or a donor node, or a wireless controller in a cloud radio access network (CRAN) scenario. Optionally, the access network device can also be a server, a wearable device, or a vehicle-mounted device, etc. For example, the access network device in vehicle to everything (V2X) technology can be a road side unit (RSU). Multiple access network devices in a communication system can be base stations of the same type or base stations of different types. The base station can communicate with the terminal directly or through a relay station. The terminal can 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 of the access network device. In the present application, the access network device is referred to as a network device. Figure 1 Figure 1 In the present application, the device for realizing the function of the network device can be the network device, or a device capable of supporting the network device to realize the function, such as a processor, a circuit, a chip, or a chip system, etc., which can be installed in the network device or used in connection with the network device. In the technical solutions provided in the present application, the device for realizing the function of the network device is taken as an example to describe the technical solutions provided in the present application.

[0086] In the present application, the device for realizing the function of the network device can be the network device, or a device capable of supporting the network device to realize the function, such as a processor, a circuit, a chip, or a chip system, etc., which can be installed in the network device or used in connection with the network device. In the technical solutions provided in the present application, the device for realizing the function of the network device is taken as an example to describe the technical solutions provided in the present application.

[0087] ​The terminal device in the present application can be a wireless terminal device capable of receiving network device scheduling and indication information. The wireless terminal device can be a device that provides voice and / or data connectivity to a user, or a handheld device with wireless connection function, or other processing devices connected to a wireless modem. For example, the terminal device can communicate with one or more core networks or the Internet through a radio access network (RAN). The terminal device can also be referred to as a terminal, user equipment (UE), mobile station, mobile terminal, etc. 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, remote medical treatment, smart power grid, smart furniture, smart office, smart wear, smart transportation, smart city, or satellite communication, etc. The terminal can be a mobile phone, tablet computer, computer with wireless transceiver function, wearable device, vehicle, aircraft (such as unmanned aerial vehicle, helicopter, airplane), hot air balloon, ship, robot, mechanical arm, or smart home device, etc. The embodiments of the present application do not limit the form of the terminal device.

[0088] In the present application, the device for realizing the function of the terminal device can be a terminal device, or a device capable of supporting the terminal device to realize the function, such as a processor, circuit, chip, chip system, etc., which can be installed in the terminal device or connected with the terminal device for use. In the technical solutions provided in the present application, the device for realizing the function of the terminal device is taken as an example to describe the technical solutions provided in the present application.

[0089] The access network device and / or terminal can be fixed or mobile. The access network device and / or terminal can be deployed on land, including indoor or outdoor, handheld or vehicle-mounted; can also be deployed on the water surface; can also be deployed on aircraft, balloons and artificial satellites in the air. The embodiments of the present application do not limit the application scenarios of the access network device and the terminal. The access network device and the terminal device can be deployed in the same scenario or different scenarios, for example, the access network device and the terminal device are deployed on land at the same time; or the access network device is deployed on land and the terminal device is deployed on the water surface, etc., which will not be listed one by one.

[0090] In actual application, a terminal can be assisted to implement wireless access by multiple network devices in cooperation, and different network devices respectively implement part of functions of a base station. For example, a network device can be a central unit (CU), a distributed unit (DU), a CU-control plane (CP), a CU-user plane (UP), or a radio unit (RU), etc. The CU and the DU can be separately arranged or can be included in the same network element, such as a baseband unit (BBU). The RU can be included in a radio frequency device or a radio frequency unit, such as a remote radio unit (RRU), an active antenna unit (AAU), or a remote radio head (RRH).

[0091] In different systems, the CU (or CU-CP and CU-UP), DU or RU can also have different names, but those skilled in the art can understand their meanings. For example, in an ORAN system, the CU can also be referred to as an O-CU (open CU), the DU can also be referred to as an O-DU, the CU-CP can also be referred to as an O-CU-CP, the CU-UP can also be referred to as an O-CU-UP, and the RU can also be referred to as an O-RU. Any of the CU (or CU-CP, CU-UP), DU and RU in this application can be implemented by a software module, a hardware module, or a combination of a software module and a hardware module. The CU (or CU-CP and CU-UP), DU and RU can implement different protocol layer functions.

[0092] To facilitate understanding of the embodiments of the present application, first, the terms and related technologies involved in the present application are briefly explained. Optionally, the explanation of some terms can also refer to the explanation in the 3rd generation partnership project (3rd generation partnership project, 3GPP) standard protocol.

[0093] 1. Model

[0094] In the embodiments of the present application, the model can include an AI model or a machine learning (ML) model. The AI model is an algorithm or computer program that can implement AI functions. The model represents the mapping relationship between the input and the output of the model. In the embodiments 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.

[0095] 2、channel state information, CSI

[0096] In a communication system (e.g., an LTE communication system or an NR communication system, etc.), a network device can decide and schedule the configurations of a terminal device's downlink data channel, such as resources, modulation and coding scheme (MCS), and precoding, based on CSI. CSI is a kind of channel information, which is an information capable of reflecting channel characteristics and channel quality.

[0097] CSI can 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), etc. Signal to interference plus noise ratio can also be referred to as signal to interference and noise ratio.

[0098] CSI can be obtained by measurement. CSI measurement refers to that a receiving end solves channel information according to a reference signal (RS) sent by a sending end, i.e., estimates channel information by using a channel estimation method. The reference signal includes, for example, channel state information reference signal (CSI-RS) or synchronization signal block (SSB), etc.

[0099] CSI can also be obtained by prediction. With the development of AI technology and ML technology, etc., in the field of communication, a model can be used to predict CSI. For example, an AI model or an ML model can be used to predict CSI at a future time based on CSI measured at a historical time.

[0100] In addition, the model can also be used for other processing of CSI, such as compression or decompression of CSI, etc.

[0101] In the embodiments of the present application, the model used for processing CSI is referred to as a CSI model.

[0102] 3. Use cases of CSI model

[0103] The network device and / or the terminal device can perform CSI-related processing through the model. In the following, several CSI model use cases are exemplarily given from two aspects of beam management (BM) and CSI feedback enhancement.

[0104] 1) Beam management

[0105] The network device and / or the terminal device can predict CSI for beam management based on the CSI model. The CSI for beam management includes, for example, the signal quality of the reference signal, such as Layer 1 reference signal received power (L1-RSRP). In the embodiments of the present application, the CSI model used for predicting the CSI for beam management can be referred to as a CSI model for beam management.

[0106] Two CSI model use cases for beam management (hereinafter referred to as beam management use cases, also referred to as BM-case) are exemplarily given.

[0107] a. Beam management use case 1: Spatial domain beam prediction based on the CSI model.

[0108] Take the application of the beam management use case 1 to the terminal device as an example, and the CSI model used by the beam management use case 1 is referred to as a CSI model 1.

[0109] Specifically, in the beam management use case 1, the terminal device can perform spatial domain downlink beam prediction on the reference signal set set A based on the CSI model 1 according to the measurement result of the reference signal set set B. Wherein, the reference signal set set A is greater than the reference signal set set B, that is, the reference signal set set A includes more reference signals than the reference signal set set B.

[0110] Optionally, the network device can send a reference signal set set B to the terminal device. The terminal device measures the layer 1 signal quality (for example, L1-RSRP) of each reference signal in the reference signal set set B. Then, the terminal device inputs the L1-RSRP of each reference signal in the reference signal set set B into the CSI model 1, and the CSI model 1 outputs the L1-RSRP of each reference signal in the reference signal set set A, or the CSI model 1 can output the probability that each reference signal in set A is the reference signal with the maximum L1-RSRP value in set A. The reference signal set set A is larger than the reference signal set set B, so that the signal quality of some reference signals (reference signals included in set A but not included in set B) is predicted through the CSI model 1, thereby realizing the prediction of some beams, that is, realizing the beam prediction in the spatial domain.

[0111] b. Beam management use case 2: beam prediction in the time domain based on the CSI model.

[0112] Continue to take the application of the terminal device in the beam management use case 2 as an example, and the CSI model used in the beam management use case 2 is referred to as the CSI model 2.

[0113] Specifically, in the beam management use case 2, the terminal device can perform time-domain downlink beam prediction on the reference signal set set C based on the CSI model 2 according to the measurement result of the reference signal set set D. The reference signal set set D includes M reference signals in the historical time period (that is, in the past). The reference signal set set C includes reference signals in the future N time points.

[0114] Optionally, the network device can send a reference signal to the terminal device. The terminal device measures the layer 1 signal quality (for example, L1-RSRP) of each reference signal in the reference signal set set D at a time point T. The reference signal set set D includes M reference signals before the time point T (that is, in the historical time period). Then, the terminal device inputs the L1-RSRP of each reference signal in the reference signal set set D into the CSI model 2, and the CSI model 2 outputs the L1-RSRP of each reference signal in the reference signal set set C, or the CSI model 2 can output the probability that each reference signal in set C is the reference signal with the maximum L1-RSRP value in set C. The reference signal set set C includes reference signals in the future N time points after the time point T. Through the CSI model 2, the signal quality of the reference signal in the future time point is predicted, thereby realizing the beam prediction in the future time point, that is, realizing the beam prediction in the time domain.

[0115] It should be noted that the L1-RSRP of each reference signal in the reference signal set D can be all or part of the input of the CSI model 2, which can be determined according to actual needs.

[0116] 2) CSI feedback enhancement

[0117] The terminal device and / or the network device can perform CSI feedback enhancement based on the CSI model, for example, predict CSI feedback information, or compress or decompress the CSI to be fed back. The CSI feedback information includes but is not limited to channel matrix, singular value vector of the channel, precoding matrix indicator (PMI), etc.

[0118] Two use cases of the CSI model for CSI feedback enhancement are exemplarily given (hereinafter referred to as CSI feedback enhancement use case).

[0119] a. CSI feedback enhancement use case 1: CSI feedback information prediction based on the CSI model.

[0120] The CSI feedback enhancement use case 1 is similar to the beam management use case 2, and is also a time domain prediction.

[0121] Taking the application of the CSI feedback enhancement use case 1 to the terminal device as an example, the CSI model used by the CSI feedback enhancement use case 1 is referred to as the CSI model 3.

[0122] Specifically, in the CSI feedback enhancement use case 1, the terminal device can predict the CSI feedback information of the future N time points based on the CSI model 3 according to the CSI measurement results of the reference signals at the M time points in the historical time period (i.e., the past).

[0123] Optionally, the network device can send the reference signals to the terminal device, and the terminal device measures the reference signals at the M time points to obtain the CSI measurement results (such as channel matrix, singular value vector of the channel, or PMI, etc.) corresponding to the M time points. Then, the terminal device inputs the CSI measurement results corresponding to the M time points into the CSI model 3, and the CSI model 3 outputs the prediction results (such as channel matrix, singular value vector of the channel, or PMI, etc.) of the CSI feedback information of the future N time points.

[0124] It should be noted that the CSI measurement results of the reference signals at the M time points can be all or part of the input of the CSI model 3, which can be determined according to actual needs.

[0125] b. CSI feedback enhancement use case 2: CSI compression based on the CSI model.

[0126] Taking the application of CSI feedback enhancement use case 2 to a terminal device as an example, and the CSI model used by the CSI feedback enhancement use case 2 is referred to as CSI model 4.

[0127] Specifically, in the CSI feedback enhancement use case 2, the terminal device can compress the CSI to be fed back into the CSI model 4, and the CSI model 4 outputs the CSI compression result. The terminal device can carry the CSI compression result in the 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 through the reconstruction CSI model to obtain the CSI fed back by the terminal device.

[0128] It should be noted that the CSI to be fed back compressed by the CSI model 4 can be a CSI measurement result measured, or a CSI prediction result predicted based on a CSI model or other ways, which is not limited. In addition, the CSI to be fed back can be all or part of the input of the CSI model 4, which can be determined according to actual needs.

[0129] It should be noted that the above AI use cases are only examples and do not represent that the AI use cases of the communication system only include these. Moreover, the terminal device can support one of the above use cases, or support multiple of the above use cases, which depends on the capability of the terminal device.

[0130] 4. Design of CSI model

[0131] The design of the CSI model mainly includes a data collection link, a model training link, a model inference link, and a model monitoring link.

[0132] Exemplarily, Figure 2 A schematic diagram of a CSI model design process is shown. As Figure 2 shown, the data collection link is used to collect data required by the model training link, the model inference link, or the model monitoring link. Taking the collection of data required by the model training link as an example, the network device can send a reference signal for model training to the terminal device, and the terminal device generates training data for model training according to the reference signal and collects a training data set.

[0133] In the model training link, the terminal device can analyze or train an initial model based on the training data set to obtain a CSI model. The CSI model represents the mapping relationship between the input and the output of the model. The CSI model is learned through model training, which is equivalent to learning the mapping relationship between the input and the output of the model using the training data.

[0134] In the model inference stage, the network device can also send a reference signal for inference to the terminal device. The terminal device generates inference data for model inference according to the reference signal. It should be noted that in the model inference stage, the inference data can form an inference data set, or can not form an inference data set, depending on the actual model requirements.

[0135] After generating the inference data, the terminal device uses the CSI model trained through the model training stage to perform inference based on the inference data, and obtains an inference result. Specifically, the inference data can be input into the CSI model, and an output of the CSI model is obtained as the inference result. In the above CSI model use case for beam management, the model inference stage is to predict the spatial domain or time domain of the CSI for beam management based on the CSI model. In the above CSI feedback enhancement use case 1, the model inference stage is to predict the CSI feedback information based on the CSI model. In the above CSI feedback enhancement use case 2, the model inference stage is to perform CSI compression based on the CSI model.

[0136] In the model monitoring stage, the network device can send a reference signal for performance monitoring to the terminal device. The terminal device generates monitoring data according to the reference signal, and collects a monitoring data set. The terminal device and / or the network device monitor the performance of the CSI model based on the monitoring data set.

[0137] There are three types of model performance monitoring:

[0138] In type 1, the terminal device can input the monitoring data set into the CSI model, and the CSI model outputs a result (referred to as a test result) after inference. Then, the terminal device performs relevant calculations and judgments according to the test result, and obtains a performance monitoring output. The performance monitoring output is, for example, whether the model meets a preset requirement. If the model meets the preset requirement, the performance monitoring output is 1, and if the model does not meet the preset requirement, the performance monitoring output is 0. The terminal device sends the performance monitoring output to the network device.

[0139] In type 2, the terminal device can input the monitoring data set into the CSI model, and the CSI model outputs a test result. Then, the terminal device sends the test result to the network device. The network device performs relevant calculations and judgments according to the test result, and obtains a performance monitoring output.

[0140] In type 3, the terminal device can input the monitoring dataset into the CSI model, and the CSI model outputs a test result. Then, the terminal device performs a relevant calculation according to the test result to determine a value of an index (referred to as a performance index) representing the performance of the CSI model. The terminal device sends the value of the performance index to the network device. The network device performs a relevant judgment according to the value of the performance index to obtain a performance monitoring output.

[0141] It should be noted that in some other embodiments, the model monitoring link can also be performed in the model inference process. Specifically, the terminal device and the network device can perform performance monitoring on the CSI model based on inference data or inference results.

[0142] The performance monitoring output of the model can be used to guide the management of the model. For example, when the performance monitoring output of the CSI model does not meet a preset requirement, it is indicated that the model is unusable, and the training dataset can be re-collected, and the model training link can be re-entered to re-train a new CSI model.

[0143] It can be understood that the network device needs to send a reference signal in the training link, the inference link, and the monitoring link of the CSI model. However, in different links, the network device can send the reference signal at different timings, with different periods, or in different ways. In embodiments of the present application, the reference signals in different links can be distinguished.

[0144] In addition, for CSI models with different functions, the network device can send different reference signals to obtain different reference signals. For example, for the CSI model in the CSI feedback enhancement use case described above, the network device can send a channel state information reference signal (CSI-RS) in the reference signal. For the CSI model in the beam management use case described above, the network device can send a CSI-RS and / or an SSB in the reference signal.

[0145] It can be understood that, in order to improve the accuracy of the model, a large amount of training data is required for the training of the CSI model. The training data is generated based on the reference signal sent by the network device. Figure 2 As can be seen from the corresponding description, the training data needs to be generated based on the reference signal sent by the network device. In the current massive MIMO scenario, the training of the CSI model requires a large amount of data of the reference signal, and therefore a large amount of air interface overhead is required for the construction of the training dataset.

[0146] Therefore, the application provides a communication method. A terminal device receives a plurality of first reference signals sent by a network device, and determines a first data set according to the plurality of first reference signals, the first data set being used for at least one of training, inference or monitoring of a first model. Then, a second data set is obtained according to the first data set, the second data set being a data set used for training a second model, the second model having the same structure as the first model. That is, the second model is a historical model corresponding to the first model, and the second data set is a historical training data set of the second model. Meanwhile, a similarity index (referred to as a first similarity index) of the second data set and the first data set satisfies a first condition, that is, the second data set is relatively similar to the first data set. The first data set can reflect a current channel state and a network scenario in which the terminal device is currently located. The second data set is relatively similar to the first data set, which indicates that a 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 a model required by 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 reduce the demand for new training data by reusing the second data set or using the experience and indication of the second model, thereby reducing the demand for reference signals sent by the network device, and further reducing air interface overhead and effectively saving network resources.

[0147] The scheme provided by the application will be described in detail below in combination with a corresponding flowchart. It can be understood that the main devices (such as terminal devices and network devices) in the illustrative flowchart are taken as examples to illustrate the execution subject of the method, but the application does not limit the execution subject of the method. For example, the devices (such as terminal devices and network devices) in the illustrative flowchart can also be chips, chip systems or processors supporting the devices to implement the method, and can also be logical modules or software capable of implementing all or part of the functions of the devices.

[0148] Here, it is uniformly stated that the messages or signaling interactions involved in the interaction flow of the embodiments of the application can adopt messages or signaling in standards or newly introduced messages or signaling, and the embodiments of the application do not make specific limitations.

[0149] Figure 3 FIG. 1 is a schematic diagram of a communication method 300 according to an embodiment of the application. It can be understood that Figure 3 The terminal device in the method 300 can be Figure 1 Any terminal device in the method 300, or a device (such as a processor, a chip or a chip system) in the terminal device. The network device can be Figure 1 Any access network device in the method 300, or a device (such as a processor, a chip or a chip system) in the access network device. As Figure 3 shown in FIG. 1, the method 300 includes the following steps:

[0150] 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.

[0151] S320, the terminal device determines a first data set based on the plurality of first reference signals; the first data set is used for at least one of training, inference or performance monitoring of the first model.

[0152] The first data set is used for at least one of training, inference or performance monitoring of the model, that is, the first data set can be a training data set of the first model, can be an inference data set of the first model, can be a monitoring data set of the first model, or can be a multi-purpose data set of the first model. For example, the first data set can be both an inference data set of the first model and a performance monitoring data set of the first model. In summary, the first data set is a data set related to the first model and can be used as an input in the training, inference or performance monitoring of the first model.

[0153] Optionally, the first model can be an AI model or an ML model. The function or type of the first model is not limited, for example, the first model can be a CSI model for beam management described above, or can be a CSI model for CSI feedback enhancement. The first model can be a single-sided model on the terminal device side, or can be a part of a terminal device-network device double-sided model on the terminal device side.

[0154] It can be understood that the function of the first model is different, and the data in the first data set can be different. For example, when the first model is a CSI model for beam management described above, the first data set can include layer 1 signal quality (such as L1-RSRP) corresponding to a plurality of reference signals; when the first model is a CSI model in the above feedback enhancement use case 1, the first data set can include CSI measurement results (such as channel matrix, singular value vector of channel or PMI, etc.) corresponding to a plurality of time points; when the first model is a CSI model in the above feedback enhancement use case 2, the first data set can include CSI to be fed back, which can be a measured CSI measurement result or a predicted CSI prediction result based on a CSI model or other means.

[0155] S330, the terminal device obtains 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 satisfies a first condition.

[0156] In the embodiments of the present application, the two models having the same structure can be understood as the initial model used for training the two models being the same. Specifically, the two models having the same structure 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 size. The network architecture being the same can include the number of network layers, the type of each layer, and the connection manner being the same, for example, both of the two models include N layers of network, each layer of network is a CNN network, and the connection manner is residual connection. The input and output being the same can include the type of input data being the same, and the type of output data being the same, for example, the input of both of the two models is a channel matrix, and the output of both of the two models is PMI. The parameter size being the same can include the number of neurons or channels of each layer of network being the same, the dimension of weight matrix being the same, and the like. It should be understood that the two models having the same structure have the same function. However, the two models having the same structure do not mean that the model parameters of the two models are the same, nor do they mean that the performance of the two models is the same. The two models having the same structure differ in training data, parameter initialization, hyperparameters, regularization, and the like, and the model parameters and performance of the two models will differ.

[0157] In the embodiments of the present application, the second data set is a data set used for training the second model, in other words, the second model is trained based on the second data set, or in other words, the second data set is the training data set corresponding to 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 in different training stages, and the two models can be considered as essentially the same initial model in different training states.

[0158] For ease of understanding, in some embodiments of the present application, the first model can also be referred to as a current model or a model to be trained, representing that the first model is a current model to be trained (or optimized). The model trained based on the initial model same as the first model in the historical time period is collectively referred to as a historical model. The second model belongs to the historical model. Correspondingly, the first data set can be referred to as a current data set, the training data set of the historical model can be collectively referred to as a historical training data set, and the second data set belongs to the historical training data set.

[0159] It can be understood that the network scenario in which the terminal device is located is different, the channel environment is different, the terminal device receives different reference signals sent by the network device, and different data sets are collected. For example, the terminal device is in an indoor scenario and an outdoor scenario, the channel environment is different, and different data sets are collected; when the terminal device is in a line of sight (LoS) scenario and a non-line of sight (NLoS) scenario, the channel environment is different, and different data sets are collected; the terminal device has different moving speeds (such as in a high-speed rail scenario and a non-high-speed rail scenario), the channel environment is different, and different data sets are collected. In other words, the data set collected by the terminal device can reflect the channel environment and the network scenario in which the terminal device is located. Further, the model trained based on different data sets can be applicable to different channel environments or different network scenarios.

[0160] In an embodiment of the present application, the similarity index of the second data set and the first data set (referred to as the first similarity index) satisfies the first condition. The first condition is a condition representing the degree of similarity between two data sets. The first similarity satisfies the first condition, which represents that the degree of similarity between the second data set and the first data set is high, that is, the second data set is more similar to the first data set. Then, it is indicated that the channel environment corresponding to the second data set is more similar to the channel environment corresponding to the first data set, and it is indicated that the network scenario corresponding to the second data set is more similar to the network scenario corresponding to the first data set. Further, the model trained based on the second data set is similar to the model trained based on the first data set, and can be applicable to the same or similar channel environment or network scenario. Then, since the second model is trained based on the second data set, the second model is more matched to the current network scenario.

[0161] The similarity index is a quantitative index for measuring the degree of similarity between two data sets. Optionally, the similarity index can be a similarity or a distance index. Among them, the similarity is positively correlated with the degree of similarity, that is, the higher the similarity, the more similar the two data sets. Optionally, the similarity can be represented by one or more of cosine similarity, intersection over union (ratio of intersection to union), Pearson correlation coefficient, normalized mean square error (NMSE), or square of generalized cosine similarity (SGCS).

[0162] The distance is negatively correlated with the degree of similarity, that is, the smaller the distance, the more similar the two data sets. Optionally, the distance can be a Euclidean distance, a Manhattan distance, or an edit distance. It can be understood that the distance and the similarity are converted to each other, for example, the similarity can be the reciprocal of the Euclidean distance.

[0163] Optionally, the first condition can be a threshold-related condition. For example, the first condition can include that the similarity index is greater than a first index threshold. Alternatively, the first condition can include that the similarity index is less than the first index threshold. The first index threshold can be predefined or configured or indicated by the network device.

[0164] Optionally, the first condition can also be a relative extreme value-related condition. For example, the first condition can include that the similarity index is the largest one among the candidate similarity indexes. Alternatively, the first condition can include that the similarity index is the smallest one among the candidate similarity indexes.

[0165] It can be understood that the specific content of the first condition can be different when different similarity indexes are used to represent the similarity degree. For example, when the similarity index is the similarity degree, the first condition can include that the similarity degree is greater than a first similarity degree threshold. Alternatively, the first condition can include that the similarity index is the largest one among the candidate similarity degrees. When the similarity index is the distance, the first condition can include that the distance is less than a first distance threshold. Alternatively, the first condition can include that the similarity index is the smallest one among the candidate distances.

[0166] In S340, the terminal device trains the first model based on the second data set and / or the second model to obtain a third model.

[0167] In an embodiment, the first model can be trained based on the second data set to obtain the third model.

[0168] Optionally, the terminal device can train the first model based on the second data set as the whole training data set to obtain the third model. For example, the second data set can be input into the first model to obtain the third model. Alternatively, the terminal device can train the first model based on the second data set as part of the training data set to obtain the third model. For example, the network device can further send a second reference signal to the terminal device, and the terminal device generates training data (which can be referred to as new training data) based on the second reference signal. The data set formed by the second data set and the new training data is used as the training data set, which is input into the first model to obtain the third model. For another example, the network device can use the data set formed by the second data set and the first data set as the training data set, which is input into the first model to obtain the third model.

[0169] As analyzed above, the second data set is relatively similar to the first data set, which indicates that the network scenario corresponding to the second data set is relatively close to the current actual network scenario. Therefore, the third model can be trained based on the second data set, and the purpose of training the first model is achieved. Moreover, in this process, the second data set is reused to train the first model, the demand for new training data is reduced, the demand for the network device to send reference signals is reduced, the air interface overhead is reduced, and network resources are saved.

[0170] In another embodiment, the terminal device can also train the first model based on the second model to obtain the third model. Optionally, the terminal device can take the second model as a teacher model and take the first model as a student model, and train the third model through knowledge distillation. This process can be understood as that the second model supervises and guides the first model, so that the first model learns the "experience" and "knowledge" of the second model, or in other words, the "experience" and "knowledge" of the second model are transferred to the first model.

[0171] As analyzed above, the second data set is relatively similar to the first data set, which indicates that the network scenario corresponding to the second data set is relatively close to the current actual network scenario. Therefore, the second model trained based on the second data set is relatively matched with the current network scenario, so that the "experience" and "knowledge" of the second model are transferred to the first model, so that the first model can also be applicable to the current network scenario, and the purpose of training the first model is achieved. Moreover, in this process, at least part of the supervised learning function is replaced by learning the "experience" and "knowledge" of the second model, the demand for new training data is reduced, the demand for the network device to send reference signals is reduced, the air interface overhead is reduced, and network resources are saved.

[0172] In yet another embodiment, the terminal device can also train the first model based on the second data set and the second model to obtain the third model. Optionally, in the process of knowledge distillation based on the second model, the second data set can be used as part or all of the training data set to train the first model. In this way, the demand for new training data is reduced from two dimensions of replacing the supervised learning function and reusing the training data, the demand for the network device to send reference signals is further reduced, the air interface overhead is further reduced, and network resources are further saved.

[0173] In summary, in the embodiments of the present application, the historical training data set (i.e., the second data set) of the second model with the same structure as the first model is obtained, and the similarity index of the second data set and the first data set currently used by the first model satisfies the first condition, i.e., the two are relatively similar. The first data set can reflect the network scenario in which the terminal device is currently 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 training effect of the model, but also reuse the historical training data set and / or replace the supervised learning function with the knowledge of the second model, thereby reducing the demand for new training data and the demand for the network device to send reference signals, and further reducing the air interface overhead and effectively saving network resources.

[0174] The application scenarios of the method provided by the present application are exemplarily described below.

[0175] In one application scenario, the communication method provided by the embodiments of the present application can be used for retraining of the model. That is, the first model is a model that has completed at least one training. In this case, the steps S310 to S340 described above can be started to be executed to retrain the first model to obtain a third model when the first model needs to be retrained.

[0176] Of course, in another application scenario, the method can also be used for initial training of the model, which is compatible with the traditional model training method. That is, the first model can be an initial model that has not been trained. In this application scenario, the steps S310 to S330 described above can be started to be executed when the first model needs to be trained. Since this is the initial training of the first model, the second data set cannot be obtained. Optionally, in the case where the second data set cannot be obtained, the terminal device can train the first model according to the traditional model training method, i.e., the network device sends the 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.

[0177] In the following embodiments of the present application, the retraining of the model is mainly taken as an example to describe the first data set for different purposes and the corresponding scenarios.

[0178] In an embodiment, the first data set is used for training the first model, or in other words, the first data set is used for training of the first model. For example, when the first model needs to be retrained, with reference to steps S310 and S320 described above, the network device can send a certain number of first reference signals used for training the first model to the terminal device, the terminal device receives the first reference signals, and generates a first data set used for training the first model according to the first reference signals. The first data set can be part or all of the training data set of the first model, and with reference to step S340 described above, the first model is trained based on the first data set, the second data set and / or the second model. In an embodiment, the training data set further includes the second data set obtained through step S330 described above. That is, the terminal device can take the first data set and the second data set as the training data set, retrain the first model, and obtain a third model.

[0179] In an embodiment, the first data set is used for inference of the first model, or in other words, the first data set is used for inference of the first model. For example, with reference to steps S310 and S320 described above, the network device can send first reference signals used for inference of the first model to the terminal device, the terminal device receives the first reference signals, and generates a first data set used for inference of the first model according to the first reference signals. The terminal device inputs the first data set into the first model for model inference. The terminal device can determine the performance of the first model according to the model inference result. When the performance of the first model is poor and needs to be retrained, the terminal device can obtain a second data set according to step S330 described above. Then, with reference to step S340 described above, the first model is retrained based on the second data set and / or the second model, and a third model is obtained.

[0180] In an embodiment, the first data set is used for performance monitoring of the first model, or in other words, the first data set is used for performance monitoring of the first model. For example, with reference to steps S310 and S320 described above, the network device can send first reference signals used for performance monitoring of the first model to the terminal device, and the terminal device receives the first reference signals sent by the network device. Figure 4 In this scenario, the communication method provided by the embodiments of the present application can include:

[0181] S310, the network device sends first reference signals used for performance monitoring of the first model to the terminal device. Correspondingly, the terminal device receives the first reference signals sent by the network device for performance monitoring of the first model.

[0182] S320, the terminal device generates a first data set used for performance monitoring of the first model according to the first reference signals.

[0183] Then, the terminal device performs performance monitoring on the first model based on the first data set. Specifically, it can include:

[0184] S410, the terminal device inputs the first data set into the first model, and the first model outputs a test result.

[0185] S420, the terminal device calculates a performance indicator of the first model based on the test result.

[0186] S430, the terminal device determines whether the performance indicator of the first model meets a third condition. If the performance indicator of the first model meets the third condition, step S440 is performed; if the performance indicator of the first model does not meet the third condition, step S330 is performed.

[0187] The third condition can be the preset requirement. The third condition is used to represent the performance of the model. If the performance indicator of the model meets the third condition, it represents that the performance of the model is good, which means that the model is available and the performance monitoring output 1 can be output. If the performance indicator of the model does not meet the third condition, it represents that the performance of the model is poor, which means that the model is unavailable and the model needs to be retrained, and the performance monitoring output 0 can be output.

[0188] Optionally, the third condition can be a threshold condition. For example, the third condition can include that the performance indicator is greater than or equal to a second indicator threshold. The type of performance indicator can be different, and the specific content of the third condition can be different. In a specific embodiment, taking the first model as a model for predicting PMI as an example, the performance indicator can be the SGCS of the PMI predicted by the first model and the measured PMI. Optionally, an SGCS threshold can be set. The third condition can include that the SGCS is greater than the SGCS threshold. If the SGCS of the first model is greater than the SGCS threshold, it represents that the performance of the first model is good, which means that the first model is available and the performance monitoring output 1 can be output. If the SGCS of the first model is less than or equal to the SGCS threshold, it represents that the performance of the first model is poor, which means that the first model is unavailable and the performance monitoring output 0 can be output.

[0189] S440, the terminal device sends the performance monitoring output 1 to the network device. Correspondingly, the network device receives the performance monitoring output 1 sent by the terminal device.

[0190] 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.

[0191] After step S440, step S310 can be performed to perform the next round of performance monitoring on the first model.

[0192] S330, the terminal device obtains a second data set.

[0193] S340, the terminal device trains the first model based on the second data set and / or the second model to obtain a third model.

[0194] In the case that the performance indicator of the first model does not meet the third condition, before or after step S330, the method can further include:

[0195] S450, the terminal device sends the performance monitoring output 0 to the network device. Correspondingly, the network device receives the performance monitoring output 0 sent by the terminal device.

[0196] The 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 the embodiment, the performance monitoring output 1 and the performance monitoring output 0 are only used as an example to illustrate different performance monitoring outputs, and do not cause any limitation on the form of the performance monitoring output.

[0197] That is, in the case of using the first data set 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. Then, according to step S340, the first model is retrained based on the second data set and / or the second model to obtain the 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 the electronic device, and improve the model training effect.

[0198] As can be seen from the above several embodiments, the communication method provided by the embodiments of the present application can be applied to the initial training scene of the model, and can also be applied to the retraining scene of the model. Moreover, the method can obtain the second data set based on the reference signal of the model training link, can also obtain the second data set based on the reference signal of the model inference link, and can also obtain the second data set based on the reference signal of the model performance monitoring link, which has wide application range and high flexibility.

[0199] Referring to Figure 5 In some embodiments, the terminal device trains the first model based on the second data set and / or the second model to obtain the third model in the above step S340, and the method can further include the following steps before the step S340:

[0200] S350, the network device sends a plurality of second reference signals to the terminal device; correspondingly, the terminal device receives the plurality of second reference signals sent by the network device.

[0201] S360, the terminal device determines a fourth data set according to the plurality of second reference signals.

[0202] The terminal device trains the first model based on the second data set and / or the second model to obtain the third model in the above step S340, which can include the following steps:

[0203] S341, the terminal device performs self-distillation on the first model based on the fourth data set and the second model to obtain the third model.

[0204] The fourth data set can be understood as a training data set corresponding to a reference signal received in real time.

[0205] In this embodiment, first, the first model is trained in a self-distillation manner based on the second model, without a large teacher model and related data of the network device, further reducing air interface overhead and saving network resources. Second, when the first model is trained based on the second model, a real-time data set (the fourth data set) is used to reflect the real channel state, so that the air interface overhead in the process of training the first model is reduced, and at the same time, the real-time real scene is also taken into account, the model training effect is improved, and the accuracy of the third model obtained by training is improved.

[0206] In some embodiments, the terminal device can set some similarity levels to represent the similarity degree. The similarity level can be used to guide the network device to send the number of real values of the second reference signal. Specifically, the following is explained:

[0207] The terminal device can divide the value of the similarity index into intervals, and each interval corresponds to a similarity level. Different similarity levels indicate different similarity degrees. The division method of the similarity level, the corresponding relationship between the similarity level and the similarity degree, etc. can be set according to actual needs in combination with the specific type of the similarity index, and the present application does not make specific limitations.

[0208] As a possible implementation manner, after obtaining the second data set, the terminal device can also determine the similarity level corresponding to the first similarity index. For ease of description, the similarity level corresponding to the first similarity index is referred to as first level information. The terminal device can send the first level information to the network device. The network device can send the corresponding number of second reference signals to the terminal device according to the similarity degree indicated by the first level information. That is, the sending quantity of the second reference signal corresponds to the first level information.

[0209] In some embodiments, the terminal device can also send the similarity level division related information to the network device, so as to synchronize the similarity degree indicated by the similarity level between the network device and the terminal device, and facilitate the network device to more accurately determine the sending quantity of the second reference signal.

[0210] Optionally, the similarity level division related information can include one or more of the following: the number of similarity levels, the interval critical value corresponding to the similarity level, or the interval width of the similarity level.

[0211] In a specific implementation, the more similar the similarity level indicated by the similarity level indication is, the fewer the second reference signals sent by the network device to the terminal device; the less similar the similarity level indicated by the similarity level indication is, the more the second reference signals sent by the network device to the terminal device. The more similar the similarity level is, 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 is, so that fewer second reference signals can be sent, thereby ensuring the model training effect while reducing the air interface overhead as much as possible and saving network resources. The less similar the similarity level is, 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 is, so that more second reference signals can be sent, thereby improving the model training effect.

[0212] In this embodiment, the amount of sending of the second reference signal is controlled based on the similarity of the first data set and the second data set, which can effectively balance the air interface overhead and the model training effect, and improve the performance of the third model obtained by training.

[0213] In addition, in some embodiments, after obtaining the second data set, the terminal device can select whether to train the first model based on the second data set according to the first similarity index. For example, 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 according to the step S340 described above. If the first similarity index does not satisfy the second condition, the first model is trained based on the reference signal sent by the network device in a traditional manner.

[0214] Optionally, the second condition can represent that the similarity level indicated by the similarity index is high. Optionally, the second condition can be a threshold-related condition. For example, the second condition can include that the similarity index is greater than a second similarity index threshold or the similarity index is less than a second similarity index threshold. Specifically, the second condition can be different according to different types of similarity indexes. For example, when the similarity index is a similarity degree, the second condition can include that the first similarity degree is greater than a second similarity degree threshold; when the similarity index is a distance (such as Euclidean distance), the second condition can include that the first Euclidean distance is less than a second distance threshold.

[0215] That is, in the case where the similarity of the second data set and the first data set is high, the first model is trained based on the second data set and / or the second model, otherwise, the first model is trained based on the reference signal sent by the network device in a traditional manner. In this way, the model training effect can be ensured, and the performance of the third model obtained by training can be improved.

[0216] The above several implementation manners are described below with the performance monitoring of the first data set for the first model as an example in combination with the accompanying drawings. Please refer to the accompanying drawings Figure 4 and Figure 6 In one implementation manner, after step S330, the method can further include:

[0217] S610, the terminal device determines first level information, the first level information representing a similarity level corresponding to the first similarity index.

[0218] The above S450, the terminal device sends the performance monitoring output 0 to the network device, including:

[0219] S451, the terminal device sends the performance monitoring output 0, the first level information and the similarity level division related information to the network device. Correspondingly, the network device receives the performance monitoring output 0 and the first level information sent by the terminal device.

[0220] After that, the method further includes:

[0221] S350, the network device sends a plurality of second reference signals to the terminal device. The sending quantity of the second reference signal corresponds to the first level information. Correspondingly, the terminal device receives a plurality of second reference signals sent by the network device.

[0222] S360, the terminal device determines a fourth data set according to the plurality of second reference signals.

[0223] S620, the terminal device determines whether the first similarity index meets a second condition; if the first similarity index meets the second condition, step S341 is executed; if the first similarity index does not meet the second condition, step S550 is executed.

[0224] 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.

[0225] S630, the terminal device trains the first model according to the fourth data set to obtain a third model.

[0226] It can be understood that step S341 and step S630 are steps executed under different first similarity indexes, so the corresponding first level information is different, the number of second reference signals sent by the network device is different, and therefore the number of data in the fourth data set in step S341 and step S630 is different.

[0227] The network device can configure or indicate a transmission manner of the information, or predefine the transmission manner of the information. The transmission manner includes signaling, a field, and the like. The terminal device transmits the information to the network device based on the transmission manner when the terminal device needs to transmit the information.

[0228] In a specific embodiment, a CSI report framework can be reused, a CSI report configuration (CSI-ReportConfig) is modified, and a new value is added in a preset field in the CSI report configuration to implement transmission of the similarity level information. In this way, existing signaling can be reused, the number of signaling types is reduced, and information transmission is simplified.

[0229] For example, the modification of the CSI report configuration can be as follows:

[0230] CSI-ReportConfig ::= {

[0231] reportConfigId = 1,

[0232] …

[0233] reportparameter = M- / / report similarity level division related information

[0234] reportQuantity = PMO- / / report performance monitoring output and similarity level

[0235] …

[0236] }

[0237] The PMO field is a field for indicating a performance monitoring output of a model. The similarity level is indicated, and specifically can be the first level information. The similarity level is added in the PMO field to implement transmission of the similarity level and reuse of the PMO field.

[0238] In the reportparameter field, M and The similarity level division related information is indicated, for example, M can be a number of similarity levels, The maximum interval threshold value can be used, and specific embodiments are described below. The reportparameter field is a field for transmitting measurement information. In this embodiment, the similarity level division related information is reported through the reportparameter field, and the reportparameter field is multiplexed.

[0239] Referring to Figure 7 In summary, the communication method provided in this embodiment includes two parts. The first part is to select a similar second data set and a second model according to a first data set, as described in steps S310 to S330 and related content. The second part is to train a first model based on the second data set and / or the second model, as described in step S340 and related content.

[0240] For ease of understanding, the first part and the second part are further described below.

[0241] 1. The first part.

[0242] First, the first similarity index is described.

[0243] In one embodiment, the first similarity index satisfies a first condition, including: the first similarity index is the one with the highest similarity degree indicated in at least one second similarity index. Each second similarity index represents the similarity degree between a third data set and the first data set, and 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 that has the same structure as the first model.

[0244] That is, the models in the terminal device that have the same structure as the first model are collectively referred to as fourth models. The fourth model can be understood as a historical model stored in the terminal device. The training data set used to train the fourth model is referred to as a third data set. The third data set can be understood as a historical training data set stored in the terminal device.

[0245] In other words, the second data set is a set selected from the historical training data sets in the terminal device. Specifically, the second data set is selected by the similarity index between each historical training data set and the first data set. The similarity index between the historical training data set and the first data set is collectively referred to as a second similarity index. The first similarity index is the one with the highest similarity degree indicated in the second similarity index, that is, the second data set is the one with the highest similarity to the first data set among the historical training data sets (i.e., third data sets) in the terminal device.

[0246] In the implementation manner, the first condition is a condition related to a relative extreme value, so that the second data set that is relatively most similar to the first data set can be screened out, and the historical training data set that can best represent the current network scene or channel environment can be screened out, so that the third model trained based on the second data set and / or the second model has a higher matching degree with the current scene, and the model training effect is improved. In addition, in the implementation manner, the relative extreme value range of the first condition is the historical training data set in the terminal device, that is, the second data set is screened out locally from the terminal device, without the need for interaction with the network device, further reducing the air interface overhead and further reducing the network resources.

[0247] In one embodiment, the similarity indicators include SGCS. Specifically, the first similarity indicator includes a first SGCS, and the first SGCS represents the SGCS 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 SGCS, and each second SGCS represents the SGCS between the feature information of one third data set and the feature information of the first data set. The first similarity indicator satisfies the first condition includes that the first SGCS is the largest one of at least one second SGCS.

[0248] Optionally, the feature information of the data set can be a feature vector of the data set. Optionally, the feature vector of the data set can be calculated based on singular value decomposition (SVD).

[0249] In another embodiment, the similarity indicators include Euclidean distance. Specifically, the first similarity indicator includes a first Euclidean distance, and 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. The second similarity indicator includes a second Euclidean distance, and each second Euclidean distance represents the Euclidean distance between the feature information of one third data set and the feature information of the first data set. The first similarity indicator satisfies the first condition includes that the first Euclidean distance is the smallest one of at least one second Euclidean distance.

[0250] In the embodiment, the first similarity indicator can be simply and quickly screened out by using the Euclidean distance, so as to determine the second data set, improve the screening efficiency, and further improve the model training efficiency.

[0251] Optionally, the feature information of the data set can be a feature distribution vector of the data set. The distribution feature vector can effectively represent the features of the set, so as to facilitate the determination of the similarity between two sets.

[0252] Optionally, the feature distribution vector of the data set can be calculated based on kernel principal component analysis (KPCA).

[0253] Next, taking the first similarity index including the first Euclidean distance as an example, the acquisition process of the second data set is further described.

[0254] Suppose that the network device sends N first reference signals to the terminal device, N being 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 . , where , represents the number of receiving antennas of the terminal device,

[0255] Suppose that the terminal device includes P fourth models, P being a positive integer. The P fourth models correspond to P third data sets respectively. The pth third data set is represented as .

[0256] 1) The terminal device calculates the distribution feature vector of the first data set based on the KPCA method .

[0257] The distribution feature vector of the first data set is represented as . Specifically, the calculation process can be as follows:

[0258] a, define the radial basis function, as formula (1):

[0259] (1).

[0260] wherein is a preset parameter, represents the local feature between two channel matrices in the data set.

[0261] For the first data set , the local feature between two original signal matrices in the first data set is .

[0262] b, the is composed into a distribution characteristic matrix , as formula (2):

[0263] (2).

[0264] c, the distribution characteristic matrix is subjected to eigenvalue decomposition, and the first P eigenvalues are taken​​​ The largest eigenvalues Construct the first dataset Distribution feature vector .

[0265] 2) The terminal device calculates the distribution feature vector of each third dataset based on the KPCA method.

[0266] The p-th third dataset The distribution feature vector is represented as .

[0267] Optionally, N samples can be sampled from each third dataset, and the distribution feature vector of the third dataset 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 dataset, thereby improving the accuracy of the similarity index calculation.

[0268] The process of calculating the distribution feature vector of the third dataset based on N samples is the same as that described above for calculating the first dataset. Distribution feature vector The process is the same, so I will not repeat it here.

[0269] The distribution feature vectors of the P third datasets can be represented as: .

[0270] 3) The terminal device calculates the distribution feature vector of each third dataset and the first dataset respectively. Distribution feature vector The Euclidean distance between them is used to obtain P second Euclidean distances.

[0271] With the p-th third dataset Distribution feature vector For example, it can be calculated according to formula (3). and Euclidean distance between :

[0272] (3)

[0273] The smaller the value, the better it represents the first dataset. With the The more similar the features of the third dataset, the better.

[0274] P second Euclidean distances can be expressed as: .

[0275] 4) Determine the smallest of the P second Euclidean distances as the first Euclidean distance, and determine the third dataset corresponding to the first Euclidean distance as the second dataset.

[0276] The first Euclidean distance is represented as , then .

[0277] The minimum one of the P second Euclidean distances is The corresponding third data set is most similar to the first data set, and thus The corresponding third data set is determined as the second data set.

[0278] In this embodiment, the distribution feature vectors of the data sets are calculated based on the KPCA method, and the similarity degree of the two data sets is determined by calculating the Euclidean distance between the two distribution feature vectors. Through this method, one of the P third data sets that is closest to the first data set can be determined as the second data set simply, quickly and effectively, so as to train the first model based on the second data set and / or the second model, improve the data acquisition efficiency, and further improve the model training efficiency.

[0279] Then, taking the first similarity index including the first Euclidean distance as an example, the division of the similarity level and the determination of the first level information are explained.

[0280] In one embodiment, M-1 interval thresholds can be pre-set, and are respectively represented as: , M is an integer greater than 1. The size relationship of the M-1 interval thresholds is: . The M-1 interval thresholds are all greater than 0. Among them, is the largest interval threshold. The M-1 interval thresholds divide M Euclidean distance intervals, which are respectively: , [ ], …, [ ], …, [ ]. Among them, , m is an integer. Optionally, the difference between adjacent two interval thresholds in the M-1 interval thresholds is equal. That is to say, the interval width of the first M-1 Euclidean distance intervals in the M Euclidean distance intervals can be equal. For example, the interval width of the first M-1 Euclidean distance intervals can be equal to .

[0281] Different Euclidean distance intervals correspond to different similarity degrees. Optionally, each Euclidean distance interval can be quantified as a quantified value, and the value represents the similarity degree of the Euclidean distance interval. The quantified value can be referred to as a similarity level. The similarity level can realize the quantification of the similarity degree by stages, simplify the similarity management, and facilitate identification and transmission. In one specific embodiment, each Euclidean distance interval can be quantified by bit bit information to obtain the corresponding similarity level. Taking the mth Euclidean distance interval [ For example, the corresponding similarity level is represented as , .

[0282] Determining the similarity level corresponding to a certain Euclidean distance is to determine the similarity level corresponding to the Euclidean distance interval to which the Euclidean distance belongs. For example, the first Euclidean distance , assuming that the Euclidean distance interval to which the first Euclidean distance belongs is , then the first Euclidean distance corresponds to the similarity level, i.e., the first level information is .

[0283] It can be understood that the lower the similarity level based on the Euclidean distance, the higher the similarity degree, and the higher the similarity level, the lower the similarity degree. That is, for , the smaller m, the higher the similarity degree, and the larger m, the lower the similarity level. The similarity level can be used to adjust the sending amount of the second reference signal used to train the first model. Specifically, after the terminal device determines the first level information, the terminal device can send the first level information to the network device, and the network device adjusts the configuration of the second reference signal resource according to the first level to control the sending amount of the second reference signal. In this way, based on the similarity degree of the second data set and the first data set, a proper amount of second reference signal is 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 the balance of the air interface overhead and the training effect.

[0284] Then, taking the first similarity index including the first Euclidean distance as an example, the second condition is further described.

[0285] In an embodiment, when the similarity index includes the Euclidean distance, the first similarity index satisfying the second condition can include that the first Euclidean distance is less than a second distance threshold. The second distance threshold can be represented as . That is, if , it means that there is no training data set similar to the current scene locally in the terminal device, and there is no historical model that matches the current scene, and then the first model is trained based on the reference signal sent by the network device in the traditional way; if , it means that there is a training data set (second data set) similar to the current scene locally in the terminal device, and there is a historical model (second model) that matches the current scene, and then the second data set is used as the training data set, and the second model is used as the self-distillation reference model, and the first model is trained by the self-distillation method.

[0286] In an embodiment, the second distance threshold may be less than or equal to the maximum interval threshold .

[0287] 2、 second part.

[0288] Here, taking 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 further described.

[0289] Optionally, the fourth data set can include input data and ground truth (hereinafter referred to as true value). The input data is used as the input of the model. The true value represents the real value or standard value of the real-time channel state, which is used as a label.

[0290] Referring to Figure 8 In an embodiment, the step S341 of self-distilling the first model based on the fourth data set and the second model to obtain the third model includes:

[0291] S810, the terminal device determines an output distribution according to the input data and the first model;

[0292] S820, the terminal device determines a soft label according to the input data and the second model;

[0293] S830, the terminal device aligns the soft label with the output distribution to obtain a soft loss;

[0294] 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;

[0295] S850, the terminal device determines a loss function according to the weighted sum of the soft loss and the hard loss;

[0296] S860, the terminal device updates the model parameters of the first model based on the loss function until the first model converges to obtain the third model.

[0297] Optionally, when self-distilling the first model, the final output of the first model can be distilled, or the process output inside the first model can be distilled.

[0298] In this embodiment, in the process of self-distilling the first model based on the second model, the soft label is used to extract the rich supervision information of the second model, and the true value is used as a hard label, which can reflect the current real channel environment and network scene. Therefore, the loss function obtained by aligning and weighting 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 scene, so as to reduce the air interface overhead while taking into account the model training effect, thereby improving the performance of the obtained third model.

[0299] In one embodiment, the output distribution includes a first probability distribution and a second probability distribution, and the output distribution is determined according to the input data and the first model, including:

[0300] a. inputting the input data into the first model to obtain a first original output;

[0301] b. performing normalization processing on the first original output by using a first temperature parameter to obtain the first probability distribution, the first temperature parameter being greater than 1;

[0302] c. performing normalization processing on the first original output by using a second temperature parameter to obtain the second probability distribution, the second temperature parameter being equal to 1.

[0303] In one embodiment, the step S820 of determining the soft label according to the input data and the second model includes:

[0304] a. inputting the input data into the second model to obtain a second original output;

[0305] b. performing normalization processing on the first original output by using a first temperature parameter to obtain a third probability distribution;

[0306] c. taking the third probability distribution as the soft label.

[0307] In the two embodiments, the normalization processing with the temperature parameter greater than 1 is performed on the first original output and the second original output respectively to realize temperature scaling of the first original output and the second original output, thereby softening the probability distribution, facilitating extraction of the generalization knowledge (or referred to as “dark knowledge”) of the second model, and enabling the first model to learn not only the result but also the correlation information between categories, improving the effectiveness of model training and the performance of the third model.

[0308] In one embodiment, the first temperature parameter is negatively correlated with the similarity degree represented by the first similarity index.

[0309] That is, the higher the similarity degree represented by the first similarity index, the smaller the first temperature parameter, and the lower the similarity degree represented by the first similarity index, the larger the first temperature parameter. The lower the similarity degree represented by the first similarity index, the less similar the second data set is to the first data set, and therefore, setting a larger first temperature parameter can realize a greater intensity of temperature scaling (i.e., increasing the temperature) to make the obtained soft label smoother, thereby being able to expose more relationships between categories in the second model, enabling the first model to learn more about the relationships between categories, improving the model learning efficiency, improving the model distillation effect, and further improving the performance of the third model.

[0310] In an embodiment, the step S830 of aligning the soft label with the output distribution to obtain the soft loss comprises: aligning the soft label with the first probability distribution to obtain the soft loss.

[0311] Optionally, the soft loss can be determined by the KL divergence between the first probability distribution and the soft label (i.e., the third probability distribution). The KL divergence is highly sensitive to the difference of non-dominant classes in the second model, and can efficiently transfer the generalization knowledge of the teacher model. Therefore, calculating the soft loss by the KL divergence can improve the model training effect.

[0312] In an embodiment, the step S840 of aligning the hard label with the output distribution to obtain the hard loss comprises: aligning the hard label with the second probability distribution to obtain the hard loss.

[0313] Optionally, the hard loss can be obtained by calculating the cross-entropy between the hard label (i.e., the true value) and the second probability distribution. By determining the hard loss by the cross-entropy, the classification target can be directly optimized, and the punishment for the wrong prediction is high. Therefore, the hard label can be anchored to prevent the second model from being misled, and the convergence speed of the model can be accelerated.

[0314] In an embodiment, the weight coefficient corresponding to the soft label is positively correlated with the similarity degree represented by the first similarity index.

[0315] That is, the higher the similarity degree represented by the first similarity index, the greater the weight coefficient corresponding to the soft label, and the lower the similarity degree represented by the first similarity index, the smaller the weight coefficient corresponding to the soft label. The higher the similarity degree represented by the first similarity index, the more similar the second data set to the first data set, which means that the second model is more suitable for the current scene, and the referenceability of the second model is stronger, or the utilization value of the second model is higher. Therefore, setting a larger weight coefficient of the soft loss can make the supervision and guidance effect of the second model on the first model account for a larger proportion. In this way, the supervision and guidance effect 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 improved.

[0316] For ease of understanding, the first model and the second model are taken as examples for the model for predicting the PMI. That is, the output of the first model and the second model is the predicted PMI. In 3GPP Release 18, the PMI can be generated based on an enhanced Type II codebook for predicted PMI and corresponding codebook parameters. The enhanced Type II codebook for predicted PMI is also referred to as a Type II Doppler-r18 codebook, etc. The PMI contains the configuration of multiple discrete parameters, such as beam indicators, wideband amplitude indicators, etc., and thus can be regarded as a classification task. The goal of the classification task can be understood as selecting the best combination of parameter values of the PMI at the predicted time from the codebook configuration.

[0317] Optionally, the original output (output logits) of the first model and the second model is defined as The original output represents the unnormalized score of each PMI option in the codebook by the first model or the second model. Wherein, represents the total number of output categories, that is, the total number of possible values of the output. For the model for predicting the PMI, that is, the total number of combinations of discrete parameters of each PMI in the Type II Doppler-r18.

[0318] For ease of distinction, the original output of the first model is referred to as the first original output, denoted as , and the original output of the second model is referred to as the second original output, denoted as . The function corresponding to the first model is denoted as . The function corresponding to the second model is denoted as . The model parameters of the first model are denoted as , and the model parameters of the second model are denoted as . The data set formed by the input data (input) in the fourth data set is denoted as In this embodiment, the input data can be a CSI matrix. The true value in the fourth data set can be denoted as The true value indicates the PMI category, for example, the correct PMI category is set to 1, and the remaining PMI categories are set to 0. The true value may also be represented by the format of the Type II Doppler-r18.

[0319] Please refer toFigure 9 The self-distillation principle diagram shown, the terminal device based on the fourth data set and the second model, the first model is self-distillation, get the third model, including the following steps:

[0320] 1), input data into the first model, get the first original output .

[0321] Specifically, the process can be represented as the following formula (4):

[0322] (4).

[0323] 2), input data into the second model, get the second original output .

[0324] Specifically, the process can be represented as the following formula (5):

[0325] (5).

[0326] 3), the first temperature parameter t is used to normalize the first original output , get the first probability distribution, the first temperature parameter t is greater than 1.

[0327] In other words, the first original output Temperature parameter greater than 1 normalization processing.

[0328] Temperature parameter greater than 1 normalization processing can also be called temperature scaling or softening processing.

[0329] In one embodiment, the first temperature parameter t can be .

[0330] 4), the second temperature parameter is used to normalize the first original output , get the second probability distribution, the second temperature parameter is 1.

[0331] In other words, the second original output Temperature parameter for 1 normalization processing. Temperature parameter for 1 normalization processing is also called standard normalization processing, get the result is standard probability distribution.

[0332] 5), the first temperature parameter t is used to normalize the second original output , get the third probability distribution.

[0333] In other words, the second original output Temperature parameter greater than 1 normalization processing.

[0334] Optionally, the first temperature parameter t can be obtained by The function is normalized, and is specifically represented as the following formula (6):

[0335] (6)

[0336] wherein, represents the cth of the C original outputs, and T represents a temperature parameter. Optionally, in the above steps 3) and 5), the first original output and the second original output are normalized using the same temperature parameter T, which is a first temperature coefficient t, i.e., T=t, t being a value greater than 1, to achieve temperature scaling and soften the distribution. In order to facilitate the distinction, the first probability distribution is denoted as , and the third probability distribution is denoted as .

[0337] In the above step 4), the second temperature coefficient used for normalizing the first original output is 1, i.e., T=1, to achieve standard normalization. In order to facilitate the description, the second probability distribution is denoted as .

[0338] The above steps 1) to 5) can also be described as follows: forward propagation generates a probability distribution, steps 1) and 2) can be described as calculating original outputs, and steps 3) and 5) can be described as temperature scaling and softening the distribution.

[0339] 6), the third probability distribution is taken as a soft label, and a soft loss of the first probability distribution and the soft label is determined.

[0340] In other words, the soft loss is determined by aligning the soft label (i.e., the third probability distribution) with the first probability distribution. The soft loss represents the confidence of the prediction result of the PMI category by the first model compared with the soft label.

[0341] In one specific embodiment, the soft loss can be determined based on the alignment of the soft label by the KL divergence. Illustratively, the calculation of the soft loss can be as follows:

[0342] (7).

[0343] wherein, represents the soft loss.

[0344] 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 learns the knowledge of the second model, or in other words, the knowledge of the second model is transferred to the first model.

[0345] 7), the true value is taken as a hard label, and a hard loss of the second probability distribution and the hard label is determined.

[0346] In other words, the hard loss is determined by aligning the hard label (i.e., the true value) with the second probability distribution (or in other words, the hard loss aligns with the hard label). The hard loss represents the confidence of the prediction result of the PMI category by the first model compared with the hard label.

[0347] In one specific embodiment, the hard loss can be determined based on the cross-entropy of aligning the hard label. Illustratively, the calculation of the hard loss can be as follows Formula (8):

[0348] (8).

[0349] wherein, represents the hard loss.

[0350] By aligning the true value, the learning of the first model is supervised by the true value, or in other words, the first model learns the true value, and the real network scene or channel environment information is learned, so that the third model obtained by training matches the real scene.

[0351] 8), a loss function is determined based on a weighted sum of the hard loss and the soft loss.

[0352] The loss function is also referred to as a total loss function, or a weighted loss function, and can be represented as . Optionally, the loss function can be determined according to the following Formula (9):

[0353] (9).

[0354] wherein, is a weight coefficient corresponding to the soft loss. is a weight coefficient corresponding to the hard loss.

[0355] In one embodiment, The value range can be In this way, the weight of the soft loss is limited within this range, preventing the soft loss from being too high or too low, which is conducive to balancing the supervision and guidance of the second model on the first model and the constraint of the true value data of the current scene on the first model, improving the model training effect, improving the matching degree of the third model obtained with the current real scene, and further improving the accuracy of the third model.

[0356] 9), based on the loss function, the model parameters of the first model are updated until the first model converges, and a third model is obtained.

[0357] Optionally, the 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 the model parameters when the first model converges, i.e., the model parameters of the third model.

[0358] For example, the following formula (10) can be used to determine the value of the soft loss:

[0359] (10).

[0360] wherein, represents the model parameters when the first model converges, i.e., the model parameters of the third model.

[0361] In one embodiment, the weight coefficient of the soft loss is which can be adjusted according to the first similarity index. Optionally, the first similarity index includes the first Euclidean distance In this case, the weight coefficient is negatively correlated with the first Euclidean distance . That is, the smaller the first Euclidean distance , the larger the weight coefficient , the larger the first Euclidean distance , and the smaller the weight coefficient .

[0362] In one specific embodiment, the relationship between the weight coefficient of the soft loss and the first Euclidean distance can be described by the following formula (11):

[0363] (11).

[0364] wherein, k is a constant used to control the decay rate of the weight coefficient as the first Euclidean distance increases.

[0365] The formula (11) can be used to quantitatively and accurately control the weight coefficient of the soft loss, thereby improving the model training effect.

[0366] In this embodiment, the smaller the first Euclidean distance , the more similar the second data set is to the first data set, which indicates that the second model is more consistent with the current scene, and the second model is more referable or has a higher utilization value. Therefore, a larger weight coefficient of the soft loss The second model can give more weight to the supervision of the first model, thereby fully utilizing the supervision of the second model, improving the model training efficiency, improving the model distillation effect, and further improving the performance of the third model.

[0367] 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 this case, the first temperature parameter t is positively correlated with the first Euclidean distance That is, the smaller the first Euclidean distance , the smaller the first temperature parameter t, and the larger the first Euclidean distance , the larger the first temperature parameter t.

[0368] In one specific embodiment, the relationship between the first temperature parameter t and the first Euclidean distance may be as follows formula (12):

[0369] (12).

[0370] The formula (12) can quantitatively and accurately control the first temperature parameter, and improve the model training effect.

[0371] In this embodiment, the larger the first Euclidean distance , the less similar the second data set is to the first data set. Therefore, setting a larger first temperature parameter t can achieve a larger intensity of temperature scaling (i.e., increasing the temperature), so that the obtained soft label is smoother, thereby exposing more relationships between classes in the second model, enabling the first model to learn more about the relationships between classes, improving the model learning efficiency, improving the model distillation effect, and further improving the performance of the third model.

[0372] It should be understood that Figures 1 to 9 The flowchart or scenario diagram shown is only for understanding, and is not intended to limit the embodiments of the present application to the examples shown in the figure. In fact, those skilled in the art can make equivalent transformations based on the examples in the Figures 1 to 9 , to obtain more implementation manners.

[0373] The above describes in detail the communication method provided by the embodiments of the present application in combination with Figures 1 to 9 The device embodiments of the present application will be described in detail below in combination with Figures 10 to 12 It should be understood that the communication device of the embodiments of the present application can perform the various communication methods of the foregoing embodiments of the present application, that is, the specific working processes of the following various products can refer to the corresponding processes in the foregoing method embodiments.

[0374] In the embodiments above, the terminal device can perform some or all of the steps in the embodiments; the network device can perform some or all of the steps in the embodiments. These steps or operations are only examples, and the embodiments of the present application can also perform other operations or variations of various operations. In addition, the various steps can be performed in different orders presented in the embodiments, and it is possible that not all operations in the embodiments of the present application are performed. Moreover, the magnitude of the serial number of each step does not mean the sequence of execution, and the execution sequence of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0375] Figure 10 is a schematic block diagram of a communication apparatus provided by the embodiments of the present application. As shown in Figure 10 , the communication apparatus 1000 can include a processing module 1010 and a communication module 1020. The processing module 1010 can implement corresponding processing functions. The communication module 1020 can implement corresponding communication functions, which can be internal communication functions of the communication apparatus 1000, or communication functions between the communication apparatus 1000 and other devices. Optionally, the communication module 1020 can also be referred to as a communication interface or a transceiver module.

[0376] Optionally, the communication apparatus 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 apparatus 1000 implements the foregoing method embodiments.

[0377] In a possible design, the communication apparatus 1000 can correspond to the terminal device in the foregoing method embodiments, or be a component (such as a circuit, a chip or a chip system, etc.) configured in the terminal device. The communication apparatus 1000 can be used to perform the steps or processes performed by the terminal device in any of the foregoing method embodiments.

[0378] For example, the communication module 1020 is configured to receive a plurality of first reference signals sent by a network device; the processing module 1010 is configured to determine a first data set based on the plurality of 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 a data set used for training a second model, the second model has the same structure as the first model, a first similarity index between the second data set and the first data set satisfies a first condition; train the first model based on the second data set and / or the second model, to obtain a third model.

[0379] In an embodiment, the first similarity indicator satisfies the first condition, including: the first similarity indicator is one with the highest similarity degree indicated in the at least one second similarity indicator; each second similarity indicator represents a similarity degree between a third data set and the first data set, each third data set is a data set used for training a fourth model, and the fourth model is a model in the terminal device with the same structure as the first model.

[0380] In an embodiment, the first similarity indicator includes a first Euclidean distance representing a Euclidean distance between the feature information of the second data set and the feature information of the first data set, and the second similarity indicator includes a second Euclidean distance representing a Euclidean distance between the feature information of a third data set and the feature information of the first data set; the first similarity indicator satisfying the first condition includes: the first Euclidean distance is the smallest one in the at least one second Euclidean distance.

[0381] In an embodiment, the feature information includes a distribution feature vector.

[0382] In an embodiment, the communication module 1020 is further configured to receive a plurality of second reference signals sent by the network device; and the processing module 1010 is further configured to determine a fourth data set according to the plurality of 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.

[0383] In an embodiment, the fourth data set includes input data and a true value; and the processing module 1010 is specifically configured to determine an output distribution according to the input data and the first model, determine a soft label according to 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, align the hard label with the output distribution to obtain a hard loss, determine a loss function based on a weighted sum of the soft loss and the hard loss, and update the first model to convergence based on the loss function to obtain the third model.

[0384] In an embodiment, in the loss function, a weight coefficient corresponding to the soft label is positively correlated with the similarity degree represented by the first similarity indicator.

[0385] In an embodiment, the processing module 1010 is specifically configured to input the input data into the first model to obtain a first original output, perform normalization processing on the first original output using a first temperature parameter to obtain a first probability distribution, the first temperature parameter is greater than 1, perform normalization processing on the first original output using a second temperature parameter 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, perform normalization processing on the second original output using the first temperature parameter to obtain a third probability distribution, and use the third probability distribution as the soft label.

[0386] In an embodiment, the first temperature parameter is negatively related to the degree of similarity represented by the first similarity index.

[0387] In an embodiment, the processing module 1010 is specifically configured to: determine a soft loss according to the KL divergence between the first probability distribution and the third probability distribution; and determine a hard loss according to the cross entropy between the second probability distribution and the true value.

[0388] In an embodiment, the communication module 1020 is further configured to: send the first grade information to the network device, the first grade information indicating the similarity grade corresponding to the first similarity index.

[0389] In an embodiment, the number of the second reference signals corresponds to the first grade information.

[0390] In an embodiment, the communication module 1020 is further configured to: send the similarity grade division related information to the network device.

[0391] In an embodiment, the similarity grade division related information includes one or more of the following: the number of the similarity grades, the interval critical value corresponding to the similarity grade, or the interval width of the similarity grade.

[0392] In an 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.

[0393] In an embodiment, the processing module 1010 is further configured to: if the first similarity index does not satisfy the second condition, receive a plurality of second reference signals sent by the network device, and train the first model based on the plurality of second reference signals to obtain a third model.

[0394] In an embodiment, the first data set is used for performance monitoring of the first model; and the processing module 1010 is specifically configured to: if the result of the performance monitoring of the first model does not satisfy a third condition, obtain a second data set.

[0395] The above is only an example, and the detailed steps or processes can refer to the description of the foregoing embodiments.

[0396] Figure 11 is a schematic block diagram of another example communication device provided by an embodiment of the present application. As shown in Figure 11 The communication device 1100 can include a processing module 1110, which can implement corresponding processing functions.

[0397] Optionally, the communication apparatus 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 apparatus 1000 implements the foregoing method embodiments.

[0398] Optionally, the communication apparatus 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 apparatus 1100, or communication functions of the communication apparatus 1100 and other apparatuses. Optionally, the communication module 1120 can also be referred to as a communication interface or a transceiver module.

[0399] In a possible design, the communication apparatus 1100 can correspond to a network device in the foregoing method embodiments, or a component (such as a circuit, a chip, or a chip system, etc.) configured in the network device. The communication apparatus 1100 can be used to execute steps or procedures performed by the network device in any of the foregoing method embodiments.

[0400] For example, the communication module 1120 is configured to: send, to a terminal device, a first reference signal; 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, inference, or performance monitoring of a first model; receive first level information sent by the terminal device; and the first level information indicates a similarity level corresponding to a first similarity index, the first similarity index characterizes a similarity degree between the first data set and a second data set, the second data set is a data set used for training a second model, the second model has a same structure as the first model, the first similarity index 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.

[0401] In one embodiment, the first similarity index satisfying the first condition includes: the first similarity index is a highest one of at least one second similarity index; each second similarity index characterizes a similarity degree between a third data set and the first data set, and each third data set is a data set used for training a fourth model, and the fourth model is a model in the terminal device that has a same structure as the first model.

[0402] In one embodiment, the first similarity index includes a first Euclidean distance, the first Euclidean distance characterizes a Euclidean distance between feature information of the second data set and feature information of the first data set, the second similarity index includes a second Euclidean distance, and each second Euclidean distance characterizes a Euclidean distance between feature information of a third data set and the feature information of the first data set; and the first similarity index satisfying the first condition includes: the first Euclidean distance is a smallest one of at least one second Euclidean distance.

[0403] In one embodiment, the feature information includes a distribution feature vector.

[0404] 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.

[0405] In one embodiment, the communication module 1120 is further configured to: receive information related to similarity level classification sent by the terminal device.

[0406] In one embodiment, the processing module 1110 is used to determine the amount of transmission of the second reference signal using first-level information.

[0407] The above are merely examples; for detailed steps or procedures, please refer to the descriptions in the foregoing embodiments.

[0408] Figure 12 This is another schematic block diagram of the communication device 1200 provided in the embodiments of this application. The communication device 1200 may be a chip, chip system, or processor, etc., in a terminal device or network device that implements the above-described methods. The communication device 1200 can be used to implement the methods described in the above-described method embodiments; for details, please refer to the descriptions in the above-described method embodiments.

[0409] like Figure 12 As shown, the communication device 1200 may include one or more processors 1210, which may also be referred to as processing units or processing modules, and can implement certain control functions. The processor 1210 may be a general-purpose processor or a dedicated processor, such as a baseband processor or a central processing unit. The baseband processor can be used to process communication protocols and communication data, while the central processing unit can be used to control the communication device 1200 (e.g., a base station, baseband chip, user, user chip), execute software programs, and process data from the software programs.

[0410] In an alternative design, the processor 1210 may also store instructions and / or data, which can be executed by the processor 1210 to cause the communication device 1200 to perform the methods described in the above method embodiments.

[0411] In another alternative 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, interface, interface circuit, or transceiver. The transceiver circuit, interface, interface circuit, or transceiver for implementing receiving and transmitting functions may be separate or integrated. The aforementioned transceiver circuit, interface, interface circuit, or transceiver may be used for reading and writing code / data, or it may be used for transmitting or relaying signals.

[0412] Optionally, one or more memories 1230 are included in the communications device 1200, which can store instructions thereon that are executable by the processor 1210 to cause the communications device 1200 to perform the methods described in the above method embodiments. Optionally, the memories 1230 can also store data. Optionally, the processor 1210 can also store instructions and / or data. The processor 1210 and the memories 1230 can be provided separately, or integrated together.

[0413] It should be understood that, in a possible design, the steps in the method embodiments provided in the present application can be completed by integrated logic circuits of hardware in the processor or instructions in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as completed by a hardware processor, or completed by a combination of hardware and software modules in the processor. The software modules can be located in random access memories, flash memories, read-only memories, programmable read-only memories or electrically erasable programmable memories, registers, or other mature storage mediums in the art. The storage medium is located in the storage, and the processor reads information in the storage medium, and combines the hardware to complete the steps of the above method. To avoid repetition, they will not be described in detail here.

[0414] In one implementation, the communications device 1200 can correspond to the terminal device in the above method embodiments, and can be used to perform the steps and / or procedures performed by the terminal device in the above method embodiments. The processor 1210 can be used to execute the instructions stored in the memories 1230, and when the processor 1210 executes the instructions stored in the memories, the processor 1210 is used to perform the steps and / or procedures of the above method embodiments corresponding to the terminal device.

[0415] In another implementation, the communications device 1200 can correspond to the network device in the above method embodiments, and can be used to perform the steps and / or procedures performed by the network device in the above method embodiments. The processor 1210 can be used to execute the instructions stored in the memories 1230, and when the processor 1210 executes the instructions stored in the memories, the processor 1210 is used to perform the steps and / or procedures of the above method embodiments corresponding to the network device.

[0416] It should be understood that the above-mentioned processing device can be one or more chips. For example, the processing device can be a field programmable gate array (FPGA), can be an application specific integrated circuit (ASIC), can also be a system on chip (SoC), can also be a central processor unit (CPU), can also be a network processor (NP), can also be a digital signal processor (DSP), can also be a micro controller unit (MCU), can also be a programmable logic device (PLD) or other integrated chip.

[0417] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically EPROM (EEPROM) or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM) and direct rambus RAM (DR RAM). It should be noted that the memory of the system and method described herein is intended to include, but not limited to, these and any other suitable types of memory.

[0418] According to the method provided in the embodiments of the present application, the present application further provides a chip system, which comprises one or more processors, and is configured to call and run instructions stored in a memory, so that the method provided in the embodiments of the present application is executed. The chip system can be composed of a chip, or can comprise a chip and other discrete devices.

[0419] The chip system can comprise input circuitry or an interface for sending information or data, and output circuitry or an interface for receiving information or data.

[0420] According to the method provided in the embodiments of the present application, the present application further provides a communication system, which comprises the network device and the terminal device described above.

[0421] According to the method provided in the embodiments of the present application, the present application further provides a computer program product, which comprises computer program codes, and when the computer program codes are run on a computer, the computer is caused to execute each step or flow of the network device and the terminal device in any of the method embodiments described above.

[0422] According to the method provided in the embodiments of the present application, the present application further provides a computer readable storage medium, which stores program codes, and when the program codes are run on a computer, the computer is caused to execute each step or flow of the network device and the terminal device in any of the method embodiments described above.

[0423] The computer readable storage medium can be the volatile memory or the non-volatile memory described above, or can comprise both the volatile memory and the non-volatile memory.

[0424] In the embodiments of the present application, each term and English abbreviation is an exemplary example given for convenience of description, and should not constitute any limitation on the present application. The present application does not exclude the possibility of defining other terms capable of achieving the same or similar functions in the existing or future protocols.

[0425] In the above embodiments, all or part of the embodiments can be realized by software, hardware, firmware or any combination thereof. When realized by software, all or part of the embodiments can be realized in the form of a computer program product. The computer program product comprises one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated.

[0426] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the division of the units is only a logical function division, and there can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0427] It should be understood that, in various embodiments of the present application, the sequence of the processes does not mean the execution sequence, and the execution sequence of the processes should be determined according to the functions and the inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0428] In summary, the above is only a preferred embodiment of the technical scheme of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A communication method, executed by a terminal device, characterized in that, include: Receive multiple first reference signals sent by network devices; A first dataset is determined based on the plurality of first reference signals; The first dataset is used for at least one of the following: training, inference, or performance monitoring of the first model; The second dataset is obtained locally from the terminal device; the second dataset is the dataset 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 dataset and the first dataset satisfies the first condition; The first model is trained based on the second dataset and / or the second model to obtain the third model.

2. The method according to claim 1, characterized in that, The first similarity index satisfies the first condition, including: The first similarity metric is the one that indicates the highest degree of similarity among at least one second similarity metric; each second similarity metric represents the degree of similarity between a third dataset and the first dataset, each of the third datasets being a dataset used to train a fourth model, the fourth model being 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 index includes a first Euclidean distance, which represents the Euclidean distance between the feature information of the second dataset and the feature information of the first dataset. The second similarity index includes a second Euclidean distance, where each second Euclidean distance represents the Euclidean distance between the feature information of the third dataset and the feature information of the first dataset. The first similarity index satisfies the 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 includes: Receive multiple second reference signals sent by network devices; The fourth dataset is determined based on the plurality of second reference signals; The step of training the first model based on the second dataset and / or the second model to obtain the third model includes: Based on the fourth dataset and the second model, the first model is self-distilled to obtain the third model.

6. The method according to claim 5, characterized in that, The fourth dataset includes input data and ground truth values; the process of self-distilling the first model based on the fourth dataset and the second model to obtain the third model includes: Based on the input data and the first model, determine the output distribution; Based on the input data and the second model, determine the soft label; Align the soft labels with the output distribution to obtain the soft loss; The true value is used as a hard label, and the output distribution is used to align the hard label to obtain the hard loss. The loss function is determined based on the weighted sum of the soft loss and the hard loss; Based on the loss function, the first model is updated until convergence, resulting in 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 index.

8. The method according to claim 6, characterized in that, 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: Input the input data into the first model to obtain the first raw output; The first original output is normalized using a first temperature parameter to obtain the first probability distribution; the first temperature parameter is greater than 1. The first original output is normalized using a second temperature parameter to obtain the second probability distribution; the second temperature parameter is equal to 1. The step of determining the soft label based on the input data and the second model includes: The input data is input into the second model to obtain the second raw output; Using the first temperature parameter, the second original output is normalized 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 the soft loss includes: The soft loss is determined based on the KL divergence between the first probability distribution and the third probability distribution; Using the true value as a hard label, and aligning the hard label with the output distribution to obtain the hard loss, including: The hard loss is determined based on the cross-entropy of the second probability distribution and the true value.

11. The method according to claim 5, characterized in that, The method further includes: Send first-level information to the network device, wherein the first-level information indicates the similarity level corresponding to the first similarity index.

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 includes: Send information related to similarity level classification to network devices.

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

15. The method according to claim 1, characterized in that, The step of training the first model based on the second dataset and / or the second model to obtain the third model includes: If the first similarity index meets the second condition, then the first model is trained based on the second dataset and / or the second model to obtain the third model.

16. The method according to claim 15, characterized in that, The method further includes: If the first similarity index does not meet the second condition, then 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 dataset is used for performance monitoring of the first model; The step of obtaining the second dataset locally from the terminal device includes: If the performance monitoring results of the first model do not meet the third condition, then the second dataset is obtained locally from the terminal device.

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

19. The method according to claim 18, characterized in that, The method further includes: Multiple second reference signals are sent 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 includes: Receive information related to similarity level classification sent by the terminal device.

21. A communication device, characterized in that, The device includes at least one processor coupled to a memory storing a program or instructions, the processor executing the program or instructions to cause the device to perform the method as described in any one of claims 1 to 20.

22. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed, they cause the computer to perform the method as described in any one of claims 1 to 20.

23. A communication system, characterized in that, Includes the communication device as described in claim 21.

24. A chip system, characterized in that, The chip system includes one or more processors, which are configured to retrieve and execute instructions stored in memory, such that the method as described in any one of claims 1 to 20 is performed.

Images