Method and apparatus in a node for wireless communication

Abstract

A method and apparatus in a node for wireless communication is provided for more sensitive and efficient performance monitoring of a model. The method in a first node for wireless communication comprises the following steps: performing performance monitoring on a first model according to a feature distribution difference between a first data feature and a second data feature; wherein the first data feature is an intermediate feature of the first model in a training stage, and the second data feature is an intermediate feature of the first model in a reasoning stage.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and more specifically, to a method and device in a node for wireless communication. Background Art

[0002] With the development of communication technology, some positioning scenarios require the introduction of artificial intelligence (AI) / machine learning (ML) models to improve positioning accuracy and computing efficiency. After deploying AI / ML models to terminal devices or network devices, how to monitor the performance of the models becomes a technical problem that needs to be solved urgently. Summary of the invention

[0003] The embodiment of the present application provides a method and an apparatus in a node for wireless communication. The following introduces various aspects involved in the present application.

[0004] In a first aspect, a method is provided for a first node for wireless communication, comprising: monitoring the performance of a first model based on a difference in feature distribution between a first data feature and a second data feature; wherein the first data feature is an intermediate feature of the first model in a training phase, and the second data feature is an intermediate feature of the first model in an inference phase.

[0005] In a second aspect, a method is provided for a second node in wireless communication, comprising: training a first model; wherein the feature distribution difference between a first data feature and a second data feature is used to monitor the performance of the first model; wherein the first data feature is an intermediate feature of the first model in a training phase, and the second data feature is an intermediate feature of the first model in an inference phase.

[0006] According to a third aspect, a first node for wireless communication is provided, comprising: a first processor, configured to monitor the performance of a first model based on a difference in feature distribution between a first data feature and a second data feature; wherein the first data feature is an intermediate feature of the first model in a training phase, and the second data feature is an intermediate feature of the first model in an inference phase.

[0007] In a fourth aspect, a second node for wireless communication is provided, comprising: a second processor for training a first model; wherein the feature distribution difference between the first data feature and the second data feature is used to monitor the performance of the first model; wherein the first data feature is an intermediate feature of the first model in the training stage, and the second data feature is an intermediate feature of the first model in the reasoning stage.

[0008] In a fifth aspect, a first node used for wireless communication is provided, comprising a transceiver, a memory and a processor, wherein the memory is used to store programs, the processor is used to call the programs in the memory and control the transceiver to receive or send signals so that the first node executes the method described in the first aspect.

[0009] In a sixth aspect, a second node used for wireless communication is provided, comprising a transceiver, a memory and a processor, wherein the memory is used to store programs, the processor is used to call the programs in the memory and control the transceiver to receive or send signals so that the second node executes the method described in the second aspect.

[0010] In a seventh aspect, an embodiment of the present application provides a communication system, which includes the first node and / or the second node described above. In another possible design, the system may also include other devices that interact with the first node or the second node in the solution provided in the embodiment of the present application.

[0011] In an eighth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program enables a computer to execute part or all of the steps in the methods of the above aspects.

[0012] In a ninth aspect, an embodiment of the present application provides a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to execute some or all of the steps in the above-mentioned various aspects of the method. In some implementations, the computer program product can be a software installation package.

[0013] In the tenth aspect, an embodiment of the present application provides a chip, which includes a memory and a processor. The processor can call and run a computer program from the memory to implement some or all of the steps described in the methods of the above aspects.

[0014] In an embodiment of the present application, through the intermediate features of the first model, the first node can determine the difference in feature distribution between the first data features and the second data features in the training phase and the inference phase (also called the deployment phase), thereby identifying whether the performance of the first model has potential degradation based on the internal features of the first model. It can be seen that the performance monitoring of the first model does not rely on the real location data and avoids the limitations of comparing input or output data. Therefore, this method can ensure that changes in model performance are discovered and responded to in a timely manner without direct location information, and the monitoring of the first model is more sensitive and effective, reducing the overall cost of system maintenance and operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The following is an example diagram of the system architecture of a wireless communication system to which the embodiments of the present application can be applied.

[0016] Figure 2 Schematic diagram of a network architecture to which the embodiments of the present application can be applied.

[0017] Figure 3A and Figure 3B The figure is a schematic diagram of the structure of a wireless protocol stack to which the embodiments of the present application can be applied.

[0018] Figure 4 Schematic diagram of neurons in a neural network to which the embodiments of the present application can be applied.

[0019] Figure 5 Schematic diagram of a neural network to which the embodiments of the present application can be applied.

[0020] Figure 6 Schematic diagram of a convolutional neural network to which the embodiments of the present application can be applied.

[0021] Figure 7 A flowchart of a method in a first node for wireless communication provided in an embodiment of the present application.

[0022] Figure 8 for Figure 7 Schematic diagram of the first intermediate layer in the illustrated method.

[0023] Fig. 9 for Figure 7 A flowchart of a possible implementation of the method shown.

[0024] Fig.10 for Figure 7 A flowchart of another possible implementation of the method is shown.

[0025] Fig.11 for Fig. 9 A flowchart of a possible implementation of the method shown.

[0026] Fig.12 for Fig.10 A flowchart of a possible implementation of the method shown.

[0027] Fig.13 A schematic diagram of the structure of a first node for wireless communication provided in an embodiment of the present application.

[0028] Fig.14 A schematic diagram of the structure of a second node for wireless communication provided in an embodiment of the present application.

[0029] Fig.15 A schematic diagram of the structure of the device provided in an embodiment of the present application.

[0030] Fig.16 A schematic diagram of the hardware modules of a communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] Communication system architecture

[0032] Figure 1 1 is a diagram showing an example of a system architecture of a wireless communication system 100 to which an embodiment of the present application can be applied. The wireless communication system 100 may include a network device 110 and a terminal device 120. The network device 110 may be a device that communicates with the terminal device 120. The network device 110 may provide communication coverage for a specific geographical area, and may communicate with the terminal device 120 located in the coverage area.

[0033] Figure 1 One network device and multiple terminal devices are exemplarily shown, for example, terminal devices 120a to 120j in the figure. Optionally, the wireless communication system 100 may include multiple network devices and each network device may include other number of terminal devices within its coverage area, which is not limited in the embodiments of the present application.

[0034] Optionally, the wireless communication system 100 may also include other network entities such as a network controller and a mobility management entity, which is not limited in the embodiments of the present application.

[0035] It should be understood that the technical solutions of the embodiments of the present application can be applied to various communication systems, such as: fifth-generation (5th-generation, 5G) system or new wireless (new radio, NR) system, long-term evolution (long term evolution, LTE) system, LTE frequency division duplex (frequency division duplex, FDD) system, LTE time division duplex (time division duplex, TDD) system, enhanced long-term evolution (advanced long term evolution, LTE-A) system, enhanced 5G (5G advanced) system, etc. The technical solutions provided in the present application can also be applied to future communication systems, such as the sixth-generation (6th-generation, 6G) mobile communication system, satellite communication system, and so on.

[0036] The terminal device in the embodiment of the present application may also be referred to as user equipment (UE), access terminal, user unit, user station, mobile station, mobile station (MS), mobile terminal (MT), remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent or user device. The terminal device in the embodiment of the present application may be a device that provides voice and / or data connectivity to a user, and may be used to connect people, objects and machines, such as a handheld device with wireless connection function, a vehicle-mounted device, etc. The terminal device in the embodiment of the present application can be a mobile phone, a tablet computer, a laptop, a PDA, a camera device, a mobile internet device (MID), a wearable device, a virtual reality (VR) device, an augmented reality (AR) device, a wireless terminal in industrial control, a wireless terminal in self-driving, a wireless terminal in remote medical surgery, a wireless terminal in smart grid, a wireless terminal in transportation safety, a wireless terminal in smart city, a wireless terminal in smart home, etc. Optionally, the terminal device can be used to act as a base station. For example, the terminal device can act as a scheduling entity, which provides sidelink signals between UEs in vehicle-to-everything (V2X) or device-to-device (D2D), etc. For example, a cellular phone and a car communicate with each other using sidelink signals. Cellular phones and smart home devices communicate with each other without relaying the communication signal through a base station.

[0037] The network device in the embodiment of the present application may be a device for communicating with a terminal device, and the network device may also be referred to as an access network device or a wireless access network device, such as a base station (BS). The network device in the embodiment of the present application may refer to a wireless access network (RAN) node or a next generation RAN (NG-RAN) node (or device) that connects a user device to a wireless network. A base station may broadly cover various names as follows, or be replaced with the following names, such as: NodeB, evolved NodeB (eNB), next generation NodeB (gNB), relay station, transmitting and receiving point (TRP), transmitting point (TP), master station (MeNB), secondary station (SeNB), multi-standard radio (MSR) node, home base station, network controller, access node, wireless node, access point (AP), transmission node, transceiver node, base band unit (BBU), remote radio unit (RRU), active antenna unit (AAU), emote radio head (RRH), central unit (CU), distributed unit (DU), positioning node, etc. A base station may be a macro base station, a micro base station, a relay node, a donor node, or the like, or a combination thereof. The base station may also refer to a communication module, modem or chip used to be set in the aforementioned equipment or device. The base station may also be a mobile switching center and a device that performs the base station function in D2D, V2X, machine-to-machine (M2M) communication, a network side device in a 6G network, and a device that performs the base station function in a future communication system. The base station may support networks with the same or different access technologies. The embodiments of the present application do not limit the specific technology and specific device form adopted by the network equipment.

[0038] Base stations can be fixed or mobile. For example, a helicopter or drone can be configured to act as a mobile base station, and one or more cells can move based on the location of the mobile base station. In other examples, a helicopter or drone can be configured to act as a device that communicates with another base station.

[0039] In some deployments, the network device in the embodiments of the present application may refer to a CU or a DU, or the network device includes a CU and a DU. The gNB may also include an AAU.

[0040] The network equipment and terminal equipment can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on the water surface; they can also be deployed on aircraft, balloons and satellites in the air. The embodiments of the present application do not limit the scenarios in which the network equipment and terminal equipment are located.

[0041] It should be understood that all or part of the functions of the communication device in the present application may also be implemented by software functions running on hardware, or by virtualization functions instantiated on a platform (eg, a cloud platform).

[0042] Figure 2 A schematic diagram of a network architecture 200 of an embodiment of the present application is exemplarily shown. The network architecture 200 illustrates the network architecture of a 5GNR / LTE / LTE-A system, which may also be referred to as a 5G system (5G system, 5GS) / evolved packet system (evolved packet system, EPS) network architecture. The network architecture 200 includes a network device 110, a terminal device 120, a 5G core network (5G core network, 5GC) / evolved packet core (evolved packet core, EPC) 210, a home subscriber server (home subscriber server, HSS) / unified data management (unified data management, UDM) 220, and at least one of an Internet service 230. Figure 2 The network device and terminal device in the figure are illustrated by taking RAN and UE as examples respectively.

[0043] like Figure 2As shown, the network device 110 provides user plane protocol and control plane protocol termination towards the terminal device 120. The network device 110 is connected to the 5GC / EPC210 via the S1 / NG interface. The 5GC / EPC210 includes a mobility management entity (MME) / authentication management field (AMF) / session management function (SMF) 211, other MME / AMF / SMF214, a service gateway (S-GW) / user plane function (UPF) 212 and a packet data network gateway (P-GW) / UPF213. MME / AMF / SMF211 is a control node that handles signaling between the terminal device 120 and the 5GC / EPC210. ​​In general, MME / AMF / SMF211 provides bearer and connection management. All user internet protocol (IP) packets are transmitted through S-GW / UPF 212, which is itself connected to P-GW / UPF 213. P-GW provides UE IP address allocation and other functions. P-GW / UPF 213 is connected to Internet services 230. Internet services 230 include operator-specific Internet protocol services, which may specifically include the Internet, intranet, IP multimedia subsystem (IMS), and packet-switched streaming services. It can be seen that network architecture 200 provides packet-switched services, but those skilled in the art will readily appreciate that the various concepts presented throughout this application can be extended to networks that provide circuit-switched services or other cellular networks.

[0044] Figure 3A and Figure 3B A schematic diagram of the wireless protocol stack structure of an embodiment of the present application is shown respectively. Figure 3A and Figure 3B The 5G wireless protocol stack is used as an example for introduction. The 5G wireless protocol stack is divided into two planes: the user plane (UP) protocol stack and the control plane (CP) protocol stack. The user plane protocol stack is the protocol cluster used for user data transmission, and the control plane protocol stack is the protocol cluster used for control signaling transmission of the 5G system. The names of the layers of each protocol stack are as follows:

[0045] like Figure 3AAs shown, the user plane protocol stack includes, from top to bottom, a service data adaptation protocol (SDAP) layer, a packet data convergence protocol (PDCP) layer, a radio link control (RLC) layer, a medium access control (MAC) layer, and a physical (PHY) layer.

[0046] like Figure 3B As shown, the control plane protocol stack includes, from top to bottom: non-access stratum (NAS); radio resource control (RRC) layer, PDCP layer, RLC layer, MAC layer and PHY layer.

[0047] It should be understood that different layers in the above protocol stack have different functions, and the communication function between the terminal device and the network device is realized through the interaction between the layers. With the development of artificial intelligence technology, artificial intelligence-assisted computing functions have penetrated into the processing implementation methods of the above protocol stack. For example, the scheduling algorithm of the MAC layer and the encoding and decoding algorithm of the PHY layer can apply artificial intelligence algorithms to improve the performance of the communication algorithm.

[0048] As an example, Figure 3A and Figure 3B The wireless protocol architecture in is applicable to the first node in this application.

[0049] As an example, Figure 3A and Figure 3B The wireless protocol architecture in is applicable to the second node in this application.

[0050] It should be understood that the interpretation of the terms in the embodiments of the present application can refer to the specification protocols TS36 series, TS37 series and TS38 series of the 3rd generation partnership project (3GPP), but can also refer to the specification protocols of the Institute of Electrical and Electronics Engineers (IEEE).

[0051] Positioning Technology

[0052] In advanced 5G and upcoming 6G network systems, positioning technology will play a key role. As an example, positioning technology needs to adapt to support a variety of application scenarios such as high accuracy, low latency and wide coverage. These scenarios include autonomous driving, Internet of Things (IoT), augmented reality (AR), smart cities and industrial automation. Traditional positioning methods mainly rely on measurement techniques such as multi-lateration, Doppler shift, angle of arrival (AOA). Doppler shift-related measurement techniques include positioning algorithms such as time difference of arrival (TDOA) and frequency difference of arrival (FDOA).

[0053] However, with the increasing complexity of 5G / 6G networks, especially in high-density user environments, non-line-of-sight (NLoS) scenarios, complex urban terrain and other conditions, traditional positioning methods require a lot of computing resources and cannot adapt to the needs of large-scale IoT and ultra-low latency applications. At this time, the introduction of AI / ML technology provides strong support for breaking through these limitations.

[0054] AI / ML technology can improve positioning accuracy by processing a large amount of nonlinear and multi-dimensional data and learning effective features and patterns from complex wireless environments. As an example, channel feature extraction based on deep learning can help improve positioning accuracy. Deep neural networks can capture the spatial and temporal characteristics of wireless signals and automatically learn the complex nonlinear changes in signal propagation. In non-line-of-sight environments, AI models can enhance their understanding of signal paths by learning complex characteristics such as signal multipath reflection and attenuation, thereby improving positioning accuracy. As an example, AI / ML technology can integrate different measurement data, such as time, angle, power and other information, and improve the estimation accuracy of position through multimodal data fusion. Especially in NLoS or environments with strong multipath effects, traditional methods may fail, and AI / ML can infer a more accurate position by fusing this information.

[0055] AI / ML technology can also help solve the problem of limited coverage. In 5G / 6G, due to the application of millimeter wave frequency bands, signals are easily blocked, resulting in limited coverage. AI / ML can expand network coverage through technologies such as predicting and compensating for blockage effects and dynamically adjusting positioning strategies. As an example, AI / ML can predict possible signal blockages by analyzing historical data and environmental information, and even continue to provide high-precision positioning through compensation technology when the signal is blocked. For example, the ML model can use the collaborative information of surrounding base stations and combine prior knowledge to predict the path and location of the signal. As an example, in a wide range of indoor and outdoor environments, the AI / ML model can dynamically adjust the positioning strategy according to changes in the current environment. For example, in dense urban areas, the AI ​​model can dynamically select the best base station or network configuration based on complex environmental changes and user density to maximize positioning accuracy and coverage.

[0056] In addition, AI / ML technology can also improve computing efficiency through intelligent data processing and resource management. The high data volume transmission of 5G / 6G networks requires higher requirements for computing efficiency. AI / ML can reduce computing resource consumption based on distributed AI models, model adaptation and compression technologies. As an example, by deploying AI / ML models at the edge of the network (such as edge computing nodes, user devices), the load on the central server can be reduced and distributed processing can be achieved. The edge AI model can process local data in real time, perform fast location estimation, and collaborate with the central model to improve overall performance when necessary. As an example, the AI / ML model can be adaptively adjusted according to network conditions and computing resources. Model compression and quantization technology can reduce computational complexity and energy consumption while ensuring positioning accuracy, making it more suitable for user devices or network environments with limited resources.

[0057] The above article introduces the support of AI / ML for positioning technology. The following article introduces various positioning methods based on AI / ML.

[0058] Direct positioning and AI-assisted positioning are two different ways of positioning technology application. The two differ in the way of achieving positioning and the degree of dependence on AI technology.

[0059] Direct positioning refers to the use of AI models to infer the location of the device directly from the raw signal data. In other words, the AI ​​model itself undertakes the entire conversion process from the input signal to the location coordinates. In direct positioning, the AI ​​model directly generates the location output by processing the raw input (such as received signal strength, arrival time, channel state information, etc.). This approach integrates signal processing and location inference into one model. Direct positioning can make full use of the powerful feature extraction capabilities of the AI ​​model, extract location-related information from complex signal patterns through technologies such as deep learning, and can generally provide higher positioning accuracy. Taking the positioning of UE, gNB and location management function (LMF) as an example, typical application scenarios of direct positioning include the following.

[0060] Scenario 1: Direct AI / ML positioning based on UE positioning and UE-side model.

[0061] Scenario 2 (Case 2b): Based on UE-assisted / LMF positioning, LMF side model is used for direct AI / ML positioning.

[0062] Scenario 3 (Case 3b): NG-RAN node assisted positioning, using LMF side model and direct AI / ML positioning.

[0063] Assisted positioning refers to the use of AI technology to enhance or optimize certain aspects of traditional positioning methods without directly inferring from signals to locations. AI in this case is usually used to improve or supplement existing positioning algorithms rather than replace them. In AI-assisted positioning, AI is usually used to improve certain aspects of the positioning system, such as improving the accuracy of signal processing, optimizing algorithm parameters, and performing error correction, while the final positioning calculation still relies on traditional methods (such as geometric algorithms, triangulation, etc.). Assisted positioning pays more attention to the optimization of existing systems and has strong adaptability. It can be fine-tuned through AI technology in different scenarios to adapt to changes in the environment.

[0064] Still taking the positioning of UE, gNB and LMF as an example, the typical application scenarios of assisted positioning include the following.

[0065] Scenario 5 (Case 2a): UE-assisted / LMF-based positioning, using UE-side model, AI / ML-assisted positioning.

[0066] Scenario 6 (Case 3a): NG-RAN node assisted positioning, using the gNB side model, AI / ML assisted positioning.

[0067] The following introduces the AI-driven network-side positioning system architecture and the UE-side positioning system architecture respectively.

[0068] In advanced 5G and upcoming 6G networks, AI-driven positioning architecture can generally be divided into two main parts: the network side (such as LMF, gNB, etc.) and the user equipment (UE) side. These two parts each play different roles in the positioning system and perform different tasks to collaboratively achieve high-precision and high-efficiency positioning services.

[0069] The network-side positioning system usually includes key network entities, such as LMF, next-generation base stations (such as gNB), core network nodes, etc. The network side is mainly responsible for centralized computing and data management, processing large-scale network information and complex operations related to positioning tasks. The network side collects a large amount of positioning-related data, which may come from multiple base stations, reference signals (such as positioning reference signals (PRS)) and UEs. The network-side nodes uniformly collect different measurement values, such as time of arrival (TOA), angle of arrival (AOA), etc.

[0070] Optionally, at the AI / ML model deployment and reasoning level, the network side can deploy AI / ML models in nodes such as LMF or gNB. These models perform positioning reasoning based on the collected measurement data. The network side has powerful computing resources that can process large amounts of input data and infer the precise location of the UE in real time.

[0071] As an example, when the AI ​​model is deployed on the LMF side, it is responsible for executing complex positioning algorithms, integrating signals from multiple measurement points, and applying deep learning or other AI technologies to improve positioning accuracy. Especially in direct AI / ML positioning scenarios, LMF is responsible for independent positioning reasoning without UE assistance.

[0072] As an example, when the AI ​​model is deployed on the gNB side, in the AI / ML-assisted positioning scenario, the model on the gNB side works together with the model on the UE side to improve positioning accuracy and coverage by optimizing signal processing and data transmission paths.

[0073] Optionally, at the signaling control level, the network side needs to perform a large number of control tasks, including coordinating the communication between the UE and the gNB, sending auxiliary data (such as location calculation auxiliary data) to the UE, managing measurement reports, etc. These control tasks ensure that the UE and the network side positioning system can collaborate efficiently.

[0074] In summary, the advantage of the network-side architecture is that it has stronger computing power and can process complex AI / ML models, which provides support for large-scale user positioning. In addition, the network side can access the global network status and topology information, and combine global data to optimize positioning reasoning.

[0075] The UE-side positioning system is mainly responsible for performing distributed positioning tasks, especially in low-latency and high-dynamic scenarios. The UE side needs to calculate in real time or assist the network side to complete the positioning task based on local signals and some network-side data. The UE side is responsible for real-time processing of local measurement data, such as signal arrival time, signal strength, angle, etc. Based on this data, the UE side can perform some pre-processing tasks and independently perform position estimation in some scenarios (such as direct positioning of the UE-side model).

[0076] Optionally, at the AI / ML model deployment and reasoning level, lightweight AI / ML models can be deployed on the UE side. These models can perform UE position estimation based on local measurement data and auxiliary data received from the network side. For example, in the direct positioning of the AI ​​model on the UE side, the AI / ML model performs reasoning based on the received signal and local data to generate a real-time position estimation result. In addition, the UE side communicates with the network side through signaling to provide measurement data or request auxiliary information, such as location calculation auxiliary data. This collaborative positioning architecture enables the UE to perform positioning reasoning more flexibly.

[0077] Compared with the network-side architecture, the advantage of the UE-side architecture is that it can perform local reasoning, reducing the delay in data uploading and processing, and is particularly suitable for application scenarios that require instant feedback (such as augmented reality, Internet of Vehicles, etc.).

[0078] Neural Networks

[0079] AI research represented by neural networks has achieved great results in many fields and will affect people's production and life for a long time in the future. A neural network can be understood as a computing model consisting of multiple interconnected neuron nodes. In a neural network, the connection strength between nodes can be expressed as a weighted value corresponding to the input signal, also known as a parameter. Each neuron node performs a weighted summation of different input signals and outputs them through a specific activation function. Neurons can rely on activation functions to achieve nonlinear mapping.

[0080] by Figure 4 The neuron shown in the figure is taken as an example. The input of the neuron can be recorded as A, and each dimension of the input is recorded as a j , the corresponding weighted value is recorded as w j , where j is 1, 2, ..., n. The input of the neuron can also set a bias term to adjust the output, such as Figure 4 The constant 1 in (corresponding weight value is denoted as b). The weight value and summation units (SU) together enhance or weaken the input. The output of SU can be input into the activation function f to obtain the output t.

[0081] Common neural networks include convolutional neural network (CNN), recurrent neural network (RNN), deep neural network (DNN), etc.

[0082] Combined with the following Figure 5 A neural network applicable to the embodiments of the present application is introduced. Figure 5 The neural network shown can be divided into three categories according to the positions of different layers: input layer 510, hidden layer 520 and output layer 530. Generally speaking, the first layer is the input layer 510, the last layer is the output layer 530, and the intermediate layers between the first layer and the last layer are all hidden layers 520.

[0083] The input layer 510 is used to input data, where the input data may be, for example, a received signal received by a receiver. The hidden layer 520 is used to process the input data, for example, to decompress the received signal. The hidden layer may also be referred to as an intermediate layer. The output layer 530 is used to output processed output data, for example, to output a decompressed signal.

[0084] See also Figure 5 , a neural network includes multiple layers, each layer includes multiple neurons. The neurons between layers can be fully connected or partially connected. For connected neurons, the output of the neurons in the previous layer can be used as the input of the neurons in the next layer.

[0085] For ease of understanding, let's take CNN as an example and combine it with Figure 6 The following is an example of multiple layers in a neural network. CNN is a deep neural network with a convolutional structure. Figure 6 As shown, the structure of CNN may include an input layer 610, a convolution layer 620, a pooling layer 630, a fully connected layer 640, and an output layer 650. The convolution layer 620, the pooling layer 630, and the fully connected layer 640 are intermediate layers of the CNN.

[0086] Each convolution layer 620 may include a plurality of convolution operators, which are also called kernels. Their function can be regarded as a filter that extracts specific information from the input signal. The convolution operator can essentially be a parameter matrix, which is usually predefined. The parameter values ​​in these parameter matrices need to be obtained through a lot of training in practical applications to help CNN make correct predictions. When CNN has multiple convolution layers, the initial convolution layer often extracts more general features, which can also be called low-level features; as the depth of CNN increases, the features extracted by the later convolution layers become more and more complex.

[0087] Pooling layer 630, in order to reduce the number of training parameters and the size of the space for extracting information, it is often necessary to periodically introduce a pooling layer after the convolution layer. The pooling layer can be introduced in a variety of ways, for example, it can be Figure 6 The one convolutional layer shown is followed by a pooling layer, or multiple convolutional layers can be followed by one or more pooling layers.

[0088] The fully connected layer 640 is used to generate the final output information. Since the convolution layer 620 and the pooling layer 630 are only responsible for extracting features and reducing the parameters brought by the input data, the processing of the convolution layer 620 and the pooling layer 630 is not enough to generate the required output information, so it is necessary to introduce the fully connected layer 640. Generally, the fully connected layer 640 can also include multiple hidden layers, and the parameters contained in the multiple hidden layers can be pre-trained according to the relevant training data of the specific task type. For example, the task type may include decoding the data signal received by the receiver. For another example, the task type may also include channel estimation based on the pilot signal received by the receiver.

[0089] After the multiple hidden layers included in the fully connected layer 640, the last layer of the entire CNN, namely the output layer 650, is used to output the result. Generally, the output layer 650 is provided with a loss function (e.g., a loss function similar to the classification cross entropy) for calculating the prediction error, or for evaluating the difference between the result output by the CNN model (also known as the predicted value) and the ideal result (also known as the true value).

[0090] It should be noted that if Figure 6 The CNN shown is only an example of a convolutional neural network. In specific applications, the convolutional neural network can also exist in the form of other network models, which is not limited in the embodiments of the present application.

[0091] Combination of the above Figures 4 to 6The various layers of the neural network and their importance in the neural network are introduced. It can be seen that the neural network model will include an input layer, an output layer, and an intermediate layer. Each layer of the neural network has its specific responsibilities and functions. The input layer is responsible for receiving and preprocessing data; the intermediate layer is responsible for feature extraction and processing of data; the output layer is responsible for converting the processed data into the required output results. Through the interaction of multiple levels, the neural network can realize the processing and analysis of complex data and extract useful information from it.

[0092] Optionally, the input or output of any intermediate layer of the neural network model can be called the intermediate feature of the model. In other words, the intermediate feature of the model can be the output of the input layer or any intermediate layer, or the input of any intermediate layer or output layer.

[0093] In the field of AI, one of the core functions of neural networks is to automatically extract features from input data. Neural networks extract features from input data through a hierarchical structure. Each layer is responsible for extracting features at different levels, from simple to complex. The feature extraction process is completed through the network's weights, biases, and nonlinear activation functions. The features extracted by neural networks include low-level features and high-level features.

[0094] Optionally, the initial layers of the network can extract low-level features of the data. For example, for images, the initial layers may extract simple features such as edges and textures. In communication signal processing, low-level features may include basic signal characteristics such as frequency, phase, amplitude, etc.

[0095] Alternatively, as the number of network layers increases, the features extracted by the neural network become more complex and abstract, and can represent higher-dimensional patterns. For example, for images, the network may recognize shapes and objects. In communications, it may extract the modulation method of complex signals, coding structures, and even patterns of interference signals.

[0096] Model Monitoring

[0097] At present, how to monitor the performance of AI / ML models is one of the topics being discussed. Taking AI models as an example, some solutions attempt to monitor by comparing the statistical characteristics of AI model inputs or outputs. These solutions involve comparing the statistical characteristics of AI model training data with the wireless signal data collected in actual deployment scenarios.

[0098] Commonly used statistical methods include kernel density estimation (KDE) algorithm and maximum mean discrepancy (MMD) algorithm, which are introduced below.

[0099] Multidimensional kernel density estimation is based on statistics of multidimensional data. In multidimensional data, assuming that the data x is d-dimensional, the sample data points can be {x1,x2,...,x n}. In multiple sample data points, each x i ∈R d , where 1≤i≤n, R d Represents d-dimensional real space. The goal of multidimensional KDE is to estimate the probability density function in the data space through the kernel function K(·). Its estimation formula is:

[0100]

[0101] Where x is the feature point to be estimated; h is the bandwidth parameter (smoothing parameter); d is the dimension of the data x; K(·) is a multidimensional kernel function, and a multidimensional Gaussian kernel function is usually selected for estimation.

[0102] Alternatively, in the multidimensional case, as a commonly used kernel function, the Gaussian kernel function is in the form of:

[0103]

[0104] Optionally, the bandwidth h is a key parameter that affects the KDE estimation effect. The larger the bandwidth, the smoother the estimated distribution; the smaller the bandwidth, the more "sharp" the estimated distribution. In multidimensional KDE, the bandwidth can be a scalar or a diagonal matrix H, which is used to select different smoothing parameters for each dimension. For each dimension j, the corresponding bandwidth parameter h j This will affect the distribution estimate along this dimension.

[0105] In summary, kernel density estimation is a non-parametric method for estimating the probability density function of a random variable. This method locally smoothes each data point in the entire data space to estimate the probability density function of the random variable. This method can provide a detailed analysis of the data distribution characteristics, especially suitable for processing high-dimensional data. Compared with MMD, KDE can more accurately describe the changes in the distribution of high-dimensional data features.

[0106] In some scenarios, after obtaining different probability density functions through the KDE method, the difference between them can be quantified using the Kullback-Leibler divergence (KL divergence, also known as relative entropy). KL divergence is an indicator that measures the difference between two probability distributions. The larger the value, the more significant the difference between the two distributions; the smaller the value, the smaller the difference.

[0107] MMD is a tool in statistics and machine learning for measuring the difference between two probability distributions. MMD is a distance metric based on kernel methods, which is used to determine whether two probability distributions are the same.

[0108] Suppose there are two sample sets X = {x1, x2, ..., x m} and Y={y1,y2,...,y n}, respectively, from two distributions P and Q. MMD can determine whether P and Q are the same by calculating the mean embedding difference of the two distributions in the reproducing kernel Hilbert space (RKHS).

[0109] The definition of MMD is based on the kernel mean embedding of samples. Its mathematical expression is as follows:

[0110]

[0111] in, is a set of functions in RKHS; f is a characteristic mapping function defined on X, represents the expected value. By introducing the kernel function k(x,y), this difference can be calculated using the kernel trick. The unbiased estimate of MMD is:

[0112]

[0113] Among them, k(x,y) is used as the kernel function, and Gaussian kernel or polynomial kernel is usually selected.

[0114] In summary, MMD is a nonparametric statistical technique of kernel method. The core of MMD is to compare the mean of the characteristic distribution of two sample sets to measure the difference between them, so as to determine whether the two sample sets are derived from the same probability distribution. The advantage of this method is that it does not require explicit estimation of the probability density function, but calculates the mean difference between samples through the kernel function, making the calculation process more efficient and direct.

[0115] The above introduces a variety of statistical methods for monitoring AI models. Regardless of the statistical method used, the scheme of monitoring by comparing the statistical characteristics of the input or output of the AI ​​model has many limitations, which are analyzed as follows. It should be noted that the embodiments of the present application are based on the following analysis, which is not a prior art and should be regarded as part of the contribution of this application to the prior art.

[0116] The statistical characteristics of the input data of an AI model usually refer to the distribution characteristics of the original data received by the AI ​​model (such as multi-dimensional channel state information (CSI), arrival time, arrival angle, etc.), such as mean, variance, etc. This type of monitoring method detects the consistency and potential deviations of the model input by comparing the changes in the distribution characteristics of the input data during the training phase and the deployment phase. There are many limitations in designing an AI model performance monitoring solution based on the differences in the statistical characteristics of the input data. Specifically, the AI ​​model performance monitoring solution based on the differences in the statistical characteristics of the input data will have the following problems.

[0117] First, the sensitivity of input data characteristics to model performance is insufficient. The statistical characteristics of input data, such as mean and variance, are low-dimensional attributes and usually only reveal the overall trend of the data. However, the performance of AI models, especially in terms of prediction accuracy, actually relies on their ability to conduct in-depth analysis of input data and extract complex features. In many cases, changes in the statistical characteristics of input data do not directly equate to corresponding changes in model performance, especially when the data changes are subtle or confined to a specific area. For example, in positioning tasks, the statistical characteristics of certain wireless signals (such as the average value of received signal strength indication (RSSI)) may have only slight fluctuations, but there is no guarantee that the deep features extracted from them by the AI ​​model are not significantly affected. For another example, model performance may be slow to respond to or difficult to detect overall changes in input data. In complex wireless environments, such as those with multipath propagation or non-line-of-sight propagation, even if the statistical characteristics of the input data appear to remain unchanged, small changes in the environment may have a significant impact on the decision-making process of the AI ​​model. These complex patterns are often not fully captured by analyzing the low-dimensional statistical characteristics of the input data alone.

[0118] Secondly, it leads to information loss. AI models usually process high-dimensional input data. For example, in the field of wireless communications, the data processed by AI models may include multi-dimensional CSI, arrival time, arrival angle, etc. When analyzing this data, statistical characteristics often only capture overall trends or basic numerical features, such as mean and variance, which may lead to ignoring the complex patterns and nuances contained in the data. In other words, when the input data is compressed into low-dimensional statistical indicators (such as mean and variance), the complex information in the original high-dimensional data (such as multipath effects and signal attenuation) may be lost, which cannot accurately reflect the internal state and potential problems of the model when processing data. In addition, this compression may also make it difficult for AI models to extract details and complex patterns of input data, because relying solely on statistical characteristics may not be able to detect subtle changes that affect the model's decision-making ability, such as interference or noise fluctuations in specific frequency bands. Therefore, in order to fully understand and optimize the performance of AI models, it is necessary to go beyond traditional low-dimensional statistical analysis and deeply explore and utilize high-dimensional information in input data.

[0119] Third, the robustness to abnormal inputs is poor. AI models may encounter abnormal situations when processing input data, such as equipment failures, malicious attacks, and carefully constructed adversarial samples. The statistical characteristics of these abnormal data may not show significant abnormalities on the surface, but they are enough to cause the AI ​​model to output significantly wrong results. The adversarial sample problem is particularly tricky because although these samples are very similar to normal inputs in statistical characteristics and will hardly trigger the alarm of the monitoring system based on statistical characteristics, they can induce the model to make extremely wrong decisions. In addition, the input data may be defective due to equipment failures or errors in the data collection process. If you rely solely on statistical characteristic monitoring, these errors may be ignored because they may not affect the overall statistical characteristics of the data.

[0120] Third, there is a lack of insight into the internal workings of the model. Monitoring based on the statistical characteristics of input data is essentially a peripheral monitoring method that only focuses on changes in data, but fails to provide in-depth insight into how the model processes the data. The limitation of this method is that it cannot reflect the internal state and operating mechanism of the AI ​​model. The AI ​​model gradually extracts and transforms features through the multi-layer structure of the neural network, and the input data is only the starting point of this complex process. Therefore, even if the statistical characteristics of the input data remain unchanged, the model may have incorrectly analyzed and processed some inputs in subsequent processing, and such internal problems are difficult to detect in time through traditional statistical monitoring methods. In addition, simply monitoring the statistical characteristics of the input data is of little help in understanding the decision-making process of the model, especially when there are complex nonlinear feature relationships between different layers of the model. This limitation may make it difficult to accurately locate and effectively solve the problem, because it is impossible to clearly determine whether the problem comes from the input data itself or a failure in the internal processing of the model.

[0121] Again, insensitivity to environmental changes. In communication scenarios, environmental changes such as multipath propagation and building obstruction can have a significant impact on wireless signals, but these impacts are often not directly reflected in the statistical characteristics of the input data. This insensitivity means that even if the communication environment undergoes changes, such as when a user enters an area with dense buildings, the statistical characteristics of the input signal may remain unchanged, while in fact the channel conditions and signal propagation paths may have changed fundamentally. In this case, the adaptability of the AI ​​model to the new environment may be reduced, but these problems may not be discovered in time based solely on the statistical characteristics of the input data. In addition, monitoring methods that rely on the characteristics of the input data cannot effectively cope with dynamic changes in scenarios, especially when the model needs to demonstrate its generalization ability in new environments. This may cause performance issues of the model in a dynamic environment to be discovered late or even not detected at all, affecting the overall performance and reliability of the model.

[0122] The statistical characteristics of the output data usually refer to the distribution of the final prediction results of the AI ​​model (such as location, classification label, etc.), and this type of monitoring method detects changes in model performance by comparing the changes in the distribution of model outputs during the training phase and the deployment phase. Although this method has a certain intuitiveness, it also has many limitations. Specifically, the AI ​​model performance monitoring scheme based on the difference in the statistical characteristics of the output data has the following problems.

[0123] First, the output data is not sensitive enough to input changes. The output data of AI models are usually low-dimensional, especially in tasks such as positioning, where the output is often a single physical quantity, such as location coordinates. This low-dimensional output does not always fully reflect the details and feature changes in the input data. Due to the lack of sensitivity of the output data to input changes, changes in its statistical characteristics are often only apparent when the input data changes significantly. For example, in a wireless signal environment, even if the multipath effect or channel conditions of the signal change subtly, these changes may not be immediately reflected in the output of the model. Therefore, if performance monitoring relies on the statistical characteristics of the output data, it may delay the discovery of abnormalities in model performance. In addition, since the input of the model is usually high-dimensional data, such as CSI, arrival time, etc., and the output is a simplified numerical value, subtle changes in the input data may be "masked" or "compressed" at the output, thereby failing to accurately reflect the true complexity of the input. This conversion from high-dimensional input to low-dimensional output may result in subtle changes in the model when processing input data not being fully reflected at the output, thereby affecting the accurate evaluation and timely monitoring of model performance.

[0124] Secondly, there is information loss in the statistical characteristics of the output data. The information loss exists mainly because the output data is the result of the input data being processed by the AI ​​model at multiple levels, and the amount of information it contains is far less rich than the original input data. In the operation of the AI ​​model, high-dimensional input data (such as signal data containing complex multipath propagation information and reflection interference) is mapped and simplified into low-dimensional output results (such as location coordinates). This conversion from high dimension to low dimension inevitably leads to information compression and loss, so that all key patterns about the environment cannot be retained in the output data. In addition, in a wireless communication environment, the impact of buildings, obstacles and other reflection factors on signal characteristics may appear as subtle changes in the input data, but these subtle changes are often not reflected in the low-dimensional output data. Therefore, relying solely on the statistical characteristics of the output data for monitoring may not capture changes in complex environments, nor can it deeply reveal whether the model has abnormalities when processing input data. This means that the output statistical characteristics can only reflect changes in the final result, but cannot provide detailed information on possible problems in the internal processing of the model.

[0125] Third, the ability to detect abnormal inputs is limited. AI models may encounter abnormal inputs during operation, including adversarial samples, data errors, or noise interference. These abnormal inputs may not immediately lead to significant changes in the output. Therefore, if monitoring relies solely on the statistical characteristics of the output data, these abnormal inputs may not be effectively detected. For example, adversarial samples are specially designed to interfere with the decision-making process of AI models without significantly changing the statistical characteristics of the input data. Adversarial samples may cause significant changes in the internal feature extraction level of the model, but since the statistical characteristics of the output data are not significantly different from normal data, it is difficult to detect such potential threats through output monitoring. In addition, the complexity of the wireless communication environment may also cause various errors in the signal transmission process, such as equipment failure or signal interference. Although these errors may be reflected at the input data level, they may not be immediately reflected in the output results. For example, in a positioning task, the location information output by the model may have only a slight deviation, while the input data may have been severely interfered. In this case, relying solely on the statistical characteristics of the output data cannot effectively identify and respond to these problems, highlighting the problem of limited ability to detect abnormal inputs.

[0126] Third, the robustness to environmental changes is poor. In communication environments, especially mobile communication and positioning scenarios, environmental changes have a significant impact on the propagation path, strength, and delay of signals. These changes may not be immediately or obviously reflected in the statistical characteristics of the model output, resulting in poor robustness of the model to environmental changes. For example, when the environment changes, such as entering an area with severe signal attenuation, the statistical characteristics of the output results may not quickly reflect this change, making it impossible for the monitoring system that relies on the output data to promptly detect the problem of reduced model adaptability. For another example, in positioning tasks, even though the wireless signal has undergone significant changes, the output coordinates may still fluctuate only within a small range. In addition, environmental changes are usually gradual, and the output statistical characteristics of the model may not show significant changes in the short term, but this does not mean that the model performance is not affected. Long-term accumulated environmental changes may cause the model performance to gradually degrade, and may eventually cause a sudden and large deviation in the output results. In this case, by the time the output data monitoring reveals the problem, the best time for intervention has often been missed.

[0127] In addition to the above problems, some solutions for real-time monitoring of the performance of AI / ML models also require real location data, that is, the benchmark truth or ground truth. However, in real-world wireless communication scenarios, due to the instability of wireless channels and the variability of environmental factors, directly obtaining real location data is both expensive and technically difficult to achieve. In other words, during the model deployment phase, the real location data of the terminal device is difficult to obtain. Therefore, in the positioning scenario of wireless communications, how to effectively monitor the performance changes of deployed AI / ML models in the absence of real location data is called a technical problem that needs to be solved.

[0128] Based on this, an embodiment of the present application proposes a method in a node for wireless communication. The method realizes real-time monitoring of the first model by comparing the feature differences of the data collected in the training phase and the inference phase, reducing the cost and complexity of obtaining the reference truth value. In addition, the data features for difference comparison are intermediate features of the first model, which overcomes the various limitations of monitoring by comparing input data or output data. Taking the AI ​​model for positioning as an example, the performance monitoring method based on extracting intermediate features from the AI ​​model solves the challenge of real-time monitoring of AI model performance in the absence of real location data in a wireless communication environment, and provides stable and reliable technical support for positioning services of wireless communications, which is particularly suitable for high-precision positioning tasks in 5G / 6G networks.

[0129] It should be understood that the embodiments of the present application can not only solve the real-time monitoring problem of the AI / ML model used for positioning, but also solve the real-time monitoring problem of the AI / ML model in other scenarios. In addition, the method of the embodiments of the present application is not limited to the UE side model, but can also be used for the NW side model, or, it can also be used for the bilateral model of the UE side and the NW side.

[0130] For ease of understanding, the following Figure 7 The method in the first node for wireless communication proposed in an embodiment of the present application is described. Figure 7 The method shown includes step S710, which may be performed by a first node.

[0131] In some implementations, the first node may be any of the communication devices described above. In some implementations, the first node may support deployment of the first model and support real-time monitoring of the first model.

[0132] As an embodiment, the first node may be any terminal device, for example, Figure 1 The terminal device 120 is shown.

[0133] As an embodiment, the first node may be a relay, such as a relay terminal.

[0134] As an embodiment, the first node may be any network device, for example, Figure 1 A network device 110 is shown.

[0135] See also Figure 7 In step S710, the performance of the first model is monitored according to the feature distribution difference between the first data feature and the second data feature.

[0136] The first model may be an AI / ML model, or a model with similar functionality. In some implementations, the first model may be used to locate the first node. In some implementations, the first model may be used to perform beam prediction.

[0137] As an embodiment, the first model is deployed on the first node side. When the first node is a terminal device, the first model is a UE side model. When the first node is a network device, the first model is a network side model. For example, in the scenario 1 described above, the first model is deployed on the UE side, and the first model can directly output the location information.

[0138] In the above embodiment, the first model is deployed on the first node side, which may mean that the first model is directly deployed on the first node, or may mean that the first model is deployed on a server connected to the first node.

[0139] In the above embodiment, the first model can be deployed on the first node side before training, or can be deployed on the first node side after training is completed. In this scenario, the inference phase in which the first node uses the first model to make predictions can also be called the deployment phase.

[0140] In some embodiments, when the first node is a terminal device, the first node can train the first model by itself, and then perform model inference and report the results. For example, the first node can collect training data or historical measurement data to train the first model. After completing the training of the first model, the first node can send the training data for model training and the trained first model to the second node.

[0141] In some embodiments, when the first node is a network device, the first node can train the first model according to the data reported by the terminal device, and then perform model inference and send the results. After completing the training of the first model, the first node can also send the training data and the trained first model to the second node.

[0142] In some embodiments, the first model can be trained by multiple devices that communicate with the first node, and then deployed to the first node side after training. For example, the second node can train the first model and send the trained first model to the first node. For another example, the server connected to the first node trains the first model and then sends the trained first model to the first node. For another example, when the first node is a terminal device, the first model can be trained by a network device that provides services for the area where the first node is located.

[0143] In the above embodiment, the server connected to the first node can send the training data of the first model to the second node.

[0144] The second node may be any network side device that communicates with the first node and supports the deployment of the first model, such as a network entity of a core network. When the first model is used for positioning, the second node may be an LMF. When the first model is used for beam prediction, the second node may be a corresponding network entity or network device.

[0145] As an embodiment, after the second node completes training of the first model, the first model can be sent to the first node. The first node receives the first model from the second node, and can also be replaced by deploying the first model on the first node side. Taking the first node as UE and the second node as LMF as an example, in the system initialization phase, the first model is first trained on the device where the LMF is located, and the training data includes historical wireless signals and their corresponding location information. After the training is completed, the first model is sent to the UE end for deployment.

[0146] As an example, LMF also bears the responsibility of model update. Once the periodic training and optimization of the first model are completed, LMF will send the updated first model to the UE. When the first model is used for positioning, the UE can use the updated model to perform the positioning task.

[0147] As an example, the UE may trigger the LMF to train or update the first model. After the training or updating is completed, the LMF transmits the new first model to the UE via signaling. The signaling is, for example, radio resource control (RRC) signaling.

[0148] In some embodiments, the second node may communicate directly with the first node or communicate with the first node through other devices. As an example, when the first node is a terminal device, the second node may communicate with the first node through a network device. For example, when the first node is a UE, the network device is a gNB, and the second node is an LMF, the gNB may act as a transit station for data transmission. In other words, the gNB is responsible for reliably forwarding the data and signaling on the LMF side to the UE, and collecting feedback information from the UE.

[0149] After the first model is deployed on the first node side, the first node can monitor the performance of the first model based on event triggering, or can periodically monitor the performance of the first model according to settings. It can be seen that the model monitoring method is used in the operation stage after the model is deployed.

[0150] In some embodiments, when the first node periodically monitors the first model, the monitoring period may be defined as T. That is, the monitoring period of the first model is T. As an embodiment, the first node may monitor the performance of the first model based on the period T. As an embodiment, the monitoring period T may be adjusted according to system settings, the complexity of the communication environment, and the service type.

[0151] The performance monitoring of the first model can be implemented based on the first data feature and the second data feature. The first data feature and the second data feature are both outputs extracted from the middle layer of the first model. The difference is that the first data feature is the middle layer output of the first model in the training phase, and the second data feature is the middle layer output of the first model in the reasoning phase. As mentioned above, the reasoning phase can also be called the deployment phase.

[0152] The intermediate layer output of the first model can be called the intermediate feature of the first model. The intermediate feature is a type of data feature. As mentioned above, the intermediate feature of the model is the data feature in the model processing process. In the embodiment of the present application, the intermediate feature of the first model is the data feature obtained after feature extraction and / or processing of any intermediate layer of the first model.

[0153] As an example, the first data feature is an intermediate feature of the first model in the training stage. The intermediate feature in the training stage can be replaced by: the intermediate layer output in the training stage, the intermediate layer feature output based on the training data, and the feature extraction of the training data by the intermediate layer.

[0154] In the above example, the training data may be stored in the second node. For example, when the second node is LMF, the training data is stored on the LMF device. The training data may include wireless signal features collected during the model training phase, providing benchmark data for subsequent feature difference calculations.

[0155] As an example, the second data feature is an intermediate feature of the first model in the inference stage. The intermediate feature in the inference stage can be replaced by: the intermediate layer output in the inference stage, the intermediate layer feature output based on the inference data, and the feature extraction of the inference data by the intermediate layer.

[0156] In the above example, the inference stage is the application stage or operation stage of the first model, and the inference data can also be called real-time data or wireless signal (channel) data of the current period. In other words, the first node can perform real-time feature collection. For example, when the first node is a terminal device, the terminal device can collect wireless signal data in the current environment in real time. Optionally, the first model can directly use this data for location prediction. At the same time, the first model on the terminal device side can also extract features from the wireless signal data to obtain second data features.

[0157] In the above example, the real-time data used as inference data may be wireless signal data randomly input during the current period, which is not limited here.

[0158] In the above example, when the first node is a UE, the UE can perform wireless communication with the network device and collect wireless signal feature data in the current environment in real time based on the communication situation. These wireless signal data can be used by the UE to extract the second data feature.

[0159] As an example, when the first node is a terminal device, it can communicate with the network device. The communication data between the first node and the network device can be used as input data of the first model to perform relevant predictions. For example, the first node can collect wireless signal data for communicating with other nodes. When the first model is used for positioning, these wireless signal data can be used as input data for positioning by the first model, and can also be used to monitor the positioning performance of the first model in real time. When the first model is used for beam prediction, these wireless signal data can be used as input data for beam prediction by the first model, and can also be used to monitor the prediction performance of the first model in real time.

[0160] In the above example, the first node can determine the wireless signal data of the current period, and input the wireless signal data into the first model to determine the second data feature corresponding to the wireless signal data of the current period. As can be seen from the foregoing, the performance monitoring of the first model can be periodic. The wireless signal data of the current period can be real-time data of the current monitoring period, and the second data feature also corresponds to the current monitoring period.

[0161] Optionally, the current time period may correspond to any monitoring period of a plurality of monitoring periods, so as to perform model monitoring based on real-time data.

[0162] Optionally, the first node may collect real-time wireless signal data based on wireless communication and perform feature extraction through the first model. For example, the first node may obtain the second data feature of the current period according to the output of the first intermediate layer.

[0163] It should be noted that the first model for obtaining the first data feature by the second node is the same as the first model for obtaining the second data feature by the first node. Once the first model is adjusted or updated, it needs to be synchronized to the first node and the second node. For example, the first model at the UE end is sent to the UE in a timely manner by the device where the LMF is located to ensure that the two models for extracting features are the same.

[0164] In some embodiments, the first data feature and the second data feature may also be obtained by the same node. For example, the first node may also directly obtain the first data feature based on the stored training data, and then obtain the second data feature based on the real-time data.

[0165] In some embodiments, the intermediate layer of the first model can be any processing layer of the first model that can perform feature extraction except the input layer and the output layer. That is, the first model can include multiple intermediate layers. For example, when the first model is an AI model, in the multi-layer neural network of the AI ​​model, each layer is responsible for extracting features at different levels. Based on the feature extraction of the intermediate layer, model monitoring can capture subtle changes in the input data, which may not be easily detected in traditional low-dimensional statistical characteristics.

[0166] The intermediate layers that extract the first data feature and the second data feature at different stages are all first intermediate layers, so as to monitor the performance changes of the first model. That is, the first data feature and the second data feature are both intermediate features output by the first intermediate layer of the first model. In some embodiments, the first intermediate layer can be one of the multiple intermediate layers included in the first model. For example, the first model includes K intermediate layers, K is a positive integer, and the first intermediate layer is one of the K intermediate layers.

[0167] As an embodiment, the first data feature is an intermediate feature output by the first intermediate layer of the first model during the training phase. During the training phase, the first model extracts features from the input data through a multi-layer neural network, and these features can reflect the complex structure and pattern of the data. For example, the first data feature can be represented as a sample set X = {x1, x2, ..., x m}, m represents the number of samples of the feature.

[0168] In some scenarios, during the training of the first model, the first model can extract the feature X using the training data and store it. In some scenarios, the first data can input the training data into the first model during monitoring to determine the first data feature.

[0169] As an embodiment, the second data feature is an intermediate feature output by the first intermediate layer of the first model in the inference stage. In the inference stage, the first model extracts features from the real-time collected data to reflect the state of the first model in the current period. For example, the first data feature can be represented as a sample set Y = {y1, y2, ..., y n}, n represents the number of samples of the feature.

[0170] In the above embodiment, the first node may collect wireless signal data in the current period or current environment in real time, and extract feature Y. In some scenarios, the first node may perform periodic data collection according to the monitoring period T.

[0171] In some embodiments, the first intermediate layer is an intermediate layer relatively close to the output layer. Features close to the output layer can be used to represent data in a deep and highly abstract manner. By comparing the high-dimensional features extracted by the first model in the training phase and the inference phase, the performance of the first model in practical applications can be more reasonably evaluated.

[0172] As an embodiment, the first intermediate layer can be directly connected to the output layer of the first model. Exemplarily, when the first model includes K intermediate layers that sequentially output features starting from the input layer, the first intermediate layer can be the Kth intermediate layer that outputs features among the K intermediate layers. In other words, the first intermediate layer can be the penultimate layer in a multi-layer network. For example, the first data feature and the second data feature are the outputs of the penultimate layer of the neural network. The output features of the penultimate layer retain important information in the input data and have a certain degree of abstractness, which is a reasonable choice for monitoring model performance.

[0173] As an embodiment, the first intermediate layer may be the second to last intermediate layer that performs feature output among the multiple intermediate layers.

[0174] For ease of understanding, the following Figure 8 An exemplary description is given. Figure 8 The model includes five processing layers for feature extraction, namely input layer 801, intermediate layers 802-804 and output layer 805. The data states processed in the five processing layers are different. The first intermediate layer can be the intermediate layer 804 close to the output layer 805, that is, the output of the intermediate layer 804 can be used as the intermediate feature of the present application.

[0175] In some embodiments, the first intermediate layer may be an intermediate layer relatively close to the input layer to increase sensitivity to input changes and reduce information loss. Figure 8 The middle layer 802 is close to the input layer 801.

[0176] Combination of the above Figure 7 and Figure 8 A method for monitoring the performance of the first model based on the data features of the intermediate layer (intermediate features) is introduced. This model monitoring method does not rely on real location data, thereby reducing costs and technical difficulties, making performance monitoring more efficient. It also improves the adaptability of the model to environmental changes, can effectively detect performance fluctuations caused by environmental changes, and make timely adjustments. Therefore, the model monitoring method proposed in the embodiment of the present application is particularly suitable for processing high-dimensional input data, such as CSI, arrival time and other data.

[0177] The following is an illustrative description of a specific method for performing model monitoring based on the first data feature and the second data feature.

[0178] In some embodiments, performance monitoring of the first model is performed based on the first data feature and the second data feature, including comparing the first data feature and the second data feature to perform performance monitoring of the first model, or calculating or post-processing the first data feature and the second data feature to perform performance monitoring of the first model.

[0179] In some embodiments, the first node directly performs performance monitoring, i.e., model monitoring, on the first model based on the feature distribution difference between the first data feature and the second data feature. For example, the first node may determine the feature distribution difference between the first data feature and the second data feature to evaluate the performance of the first model. The performance evaluation result determined according to the feature distribution difference is used to complete the model monitoring. The method of model monitoring through feature distribution difference may also be referred to as a performance monitoring method based on feature distribution difference, i.e., a feature distribution difference monitoring method. In this method, when the performance of the first model exceeds a preset state, the first node or system may determine that there is a problem with the first performance based on the feature distribution difference, thereby triggering a correction (also called adjustment) mechanism or update mechanism of the first model.

[0180] As an embodiment, the update mechanism may refer to updating the first model. Optionally, the update mechanism of the first model may implement an overall update of the first model. Optionally, when the first model is divided into multiple parts, the update mechanism may also be an update of a certain part. For example, the node may update the current AI model according to the model monitoring result to obtain an updated AI model.

[0181] As an embodiment, the correction mechanism may refer to adjusting the first model. Optionally, the correction mechanism of the first model may implement a local correction of the first model. The local correction may include parameter adjustment or model adjustment, which is not limited here. For example, the node may adjust certain parameters of the first model according to environmental changes to obtain a corrected AI model.

[0182] In some embodiments, the first node may determine whether to trigger the update or correction of the first model based on the feature distribution difference and a preset threshold. The preset threshold may be a preset error threshold. The preset threshold may be predefined or determined based on higher layer signaling.

[0183] As an embodiment, when the feature distribution difference is greater than or equal to a first preset threshold, the update mechanism of the first model is triggered; or, when the feature distribution difference is greater than or equal to a second preset threshold, the correction mechanism of the first model is triggered.

[0184] As an embodiment, when the feature distribution difference is less than a first preset threshold, the update mechanism of the first model is not triggered; or, when the feature distribution difference is less than a second preset threshold, the correction mechanism of the first model is not triggered.

[0185] As an embodiment, when the feature distribution difference is greater than a first preset threshold, the update mechanism of the first model is triggered; or, when the feature distribution difference is greater than a second preset threshold, the correction mechanism of the first model is triggered.

[0186] As an embodiment, when the feature distribution difference is less than or equal to the first preset threshold, the update mechanism of the first model is not triggered; or, when the feature distribution difference is less than the second preset threshold, the correction mechanism of the first model is not triggered.

[0187] In the above embodiment, the first preset threshold is the same as the second preset threshold; or, the first preset threshold is different from the second preset threshold.

[0188] In the above-mentioned embodiment, both the updating mechanism and the correction mechanism are for optimizing the first model to improve the model accuracy.

[0189] In the above embodiment, when the update mechanism is triggered, the second node can receive the update request of the first model triggered by the first node. When the correction mechanism is triggered, the second node can receive the correction request of the first model triggered by the first node.

[0190] In some embodiments, the feature distribution difference may include at least one of the following: a data distribution error of the first data feature and the second data feature determined by a maximum mean difference (MMD) algorithm; a KL divergence of the first data feature and the second data feature determined by a kernel density estimation (KDE) algorithm. For example, the first node may determine the data distribution error, i.e., the feature distribution difference, by a feature difference calculation scheme based on MMD. As another example, the first node may determine the KL divergence, i.e., the feature distribution difference, by a feature distribution difference calculation scheme based on KDE. The KL divergence is as described above. As another example, the first node may determine the data distribution error and the KL divergence by the MMD algorithm and the KDE algorithm, respectively, in different time periods or the same time period.

[0191] As an embodiment, the data distribution error is related to the calculation result of the MMD algorithm. For example, the data distribution error can represent the calculation result of the MMD algorithm. For example, the data distribution error can be an MMD value obtained by the MMD algorithm. For another example, the data distribution error can be determined according to the MMD value obtained by the MMD algorithm. For simplicity, the calculation result of the MMD algorithm is simplified to the MMD result in the following text.

[0192] As an embodiment, when the feature distribution difference is a data distribution error between the first data feature and the second data feature determined by the MMD algorithm, the first preset threshold or the second preset threshold may be a preset error threshold.

[0193] As an embodiment, when the feature distribution difference is the KL divergence of the first data feature and the second data feature determined by the KDE algorithm, the first preset threshold or the second preset threshold may be a threshold preset for the KL divergence.

[0194] In the above embodiment, the data distribution error determined by the MMD algorithm can measure the difference in probability distribution between the first data feature and the second data feature; the KL divergence determined by the KDE algorithm can indicate the difference in probability distribution between the first data feature and the second data feature.

[0195] Taking the first model used for positioning as an example, the following introduces the calculation solutions based on MMD and KDE respectively in combination with two embodiments.

[0196] Example 1: Feature difference calculation solution based on MMD

[0197] In this embodiment, the feature distribution difference between the first data feature and the second data feature is the data distribution error. Through the feature difference calculation and monitoring mechanism based on MMD, the system can flexibly and efficiently evaluate the performance change of the first model, reduce the cost of relying on real data, and ensure the positioning accuracy of the model in complex wireless environments.

[0198] In this embodiment, the MMD algorithm can measure the difference between two distributions based on a kernel function (usually a Gaussian kernel).

[0199] As an embodiment, the first data feature determined by the training data can be expressed as X={x1, x2, ..., x m}, extracted by the first model on the second node and stored in the first node or the device where the first node is located. The second data feature determined by the real-time collected wireless signal data can be expressed as Y = {y1, y2, ..., y n}, extracted from the first model at the first node end and stored at the first node end.

[0200] For the first data feature and the second data feature mentioned above, the calculation formula of the MMD result is as follows:

[0201]

[0202] Among them, k(x,y) is the kernel function. The kernel function can use the Gaussian kernel function, and the formula is as follows:

[0203]

[0204] Among them, σ is the bandwidth parameter of the kernel function, which controls the measurement scale of the distance in the feature space; Indicates expected value.

[0205] Through this formula, the MMD result can measure the mean difference between the distribution of the first data feature and the distribution of the second data feature in the kernel function mapping space. If the MMD result is small, it means that the two feature distributions are very similar, and the performance of the first model in the new environment is close to the training stage; if the MMD result is large, it indicates that there is a significant deviation in the distribution of the two features, and the first model may need to be adjusted or updated.

[0206] In actual calculations, the MMD result is approximated by a finite set of samples. The discrete form of the MMD result can be:

[0207]

[0208] Where m represents the number of samples of features extracted from the training data; n represents the number of samples of features of wireless signals collected in real time within a model performance monitoring period T; k(x i ,x j )、k(y i ,y j ) represents the calculation result of the kernel function between feature samples.

[0209] Through the discrete form of formula ①, the system can quickly calculate the MMD value (i.e., data distribution error) to measure the difference in feature distribution between the data features of the training data and the data features in the new environment / current environment.

[0210] In this embodiment, the data distribution error is used to compare the distribution of training data features stored in the LMF (second node) with the distribution of wireless signal features collected in real time by the UE (first node). Through this comparison, we can detect whether changes in the deployment environment have affected the performance of the first model. If the feature distribution difference obtained by the MMD algorithm exceeds a preset threshold (a first preset threshold or a second preset threshold), this indicates that the first model may need to be adjusted or optimized to adapt to new environmental conditions.

[0211] For ease of understanding, the following Fig. 9 The process of the first embodiment is schematically described. Fig. 9 It is introduced from the perspective of the interaction between the first node and the second node. Fig. 9 The monitoring in can be periodic monitoring to optimize the first model.

[0212] See also Fig. 9 In step S910, the first node sends the second data feature to the second node. The first node may first determine the second data feature. Optionally, the first node may periodically collect wireless signal data in real time based on the setting of the monitoring period and determine the second data feature. In this embodiment, the first node may upload the second data feature to the second node via signaling for subsequent feature difference calculation.

[0213] In step S920, the second node determines the data distribution error according to the first data feature and the second data feature. The second node can determine the first data feature according to the training data, and then determine the data distribution error according to formula ①. When the second node is LMF, the LMF end can receive the second data feature from the first node, and combine it with the first data feature determined according to the training data, and then use the MMD algorithm to calculate the difference between the two feature distributions. For example, LMF can calculate formula (1) (kernel mean difference between real-time acquisition features), formula (2) (kernel mean difference between training data features) and formula (3) (kernel mean difference between training data and real-time data) in sequence according to formula ①, and finally obtain the calculation result of the MMD algorithm.

[0214] In step S930, the first node receives the data distribution error from the second node. After the MMD calculation is completed, the second node transmits the calculated MMD result (data distribution error) back to the first node via signaling. This process also generates additional signaling overhead.

[0215] After receiving the data distribution error, the first node compares it with a preset threshold (the first preset threshold or the second preset threshold). If the data distribution error is less than the preset threshold, it means that the feature distribution of the first model in the current environment is generally consistent with the training data features, and the model performance has no obvious deviation. If the data distribution error exceeds the threshold, it indicates that the model performance may be affected by environmental changes, and the system can trigger an update or other correction mechanism for the first model.

[0216] It should be noted that when there is no new model update, the first data feature determination on the LMF side and the calculation of formula (2) in formula ① only need to be performed once until the next update. In other words, when the model performance is normal, some calculations in formula ① only need to be performed once until the first model is updated. This mechanism effectively reduces the computational burden and signaling overhead.

[0217] In the first embodiment, the additional signaling overhead is mainly reflected in the two aspects of feature transmission and MMD result transmission. In feature transmission, the first node needs to regularly send the second data feature of the data collected in the current period to the second node. In the MMD result transmission, the second node transmits the MMD result back to the first node after the calculation is completed, so that the first node can make a performance judgment of the first model. The first embodiment optimizes the signaling overhead. When the first node is a UE and the second node is a LMF, the calculation process of the MMD algorithm is mainly performed on the LMF side, and the UE side only needs to send the collected data features to the LMF side through signaling. This design reduces the computational burden on the UE side and reduces the overall signaling overhead. In particular, when the data distribution error does not trigger the threshold, the frequency of data transmission can be reduced, thereby further saving network resources.

[0218] In addition, embodiment 1 can realize periodic monitoring and quickly respond to environmental changes. The system regularly performs monitoring tasks, collects wireless signal feature data in the current environment of the UE, and uses the MMD algorithm to analyze the distribution differences between these data and the feature data in the training phase. Once the data distribution error determined by the MMD algorithm exceeds the set threshold, this indicates that the performance of the model in the new environment may be deviated, which may affect the accuracy of positioning. In view of the complexity and ever-changing nature of the wireless communication environment, the MMD algorithm, as a detection tool, can quickly capture subtle changes in the environment. This monitoring method based on feature differences can issue early warnings in a timely manner to ensure that the first model can be adjusted accordingly according to actual changes in the environment, thereby maintaining the accuracy of positioning and the stability of the system.

[0219] Example 2: Feature distribution difference calculation solution based on KDE

[0220] In this embodiment, the feature distribution difference between the first data feature and the second data feature is KL divergence. KDE can be used to accurately estimate the data feature distribution in the training phase and the deployment phase. By comparing the distribution difference of the two sets of features, the performance of the first model can be evaluated.

[0221] Similar to the first embodiment, the feature distribution of the training data is estimated from the training data stored on the second node, and the first node collects the wireless signal data in real time and estimates its feature distribution. The difference is that after obtaining the two sets of feature distributions, the second embodiment uses KL divergence to quantify the difference between the two. When the KL divergence exceeds the set threshold, it may mean that the model deployment environment has changed, resulting in a significant shift in the input feature distribution of the model. In this case, the model may need to be adjusted or optimized to adapt to the new environmental conditions.

[0222] As an embodiment, the first node may receive the probability density distribution of the first data feature from the second node. After determining the probability density distribution of the second data feature, the first node may calculate the KL divergence based on the two probability density distributions to obtain the feature distribution difference between the first data feature and the second data feature. Optionally, the probability density distribution may be a list formed by multiple kernel density estimation values.

[0223] In the above embodiment, the first node may receive a plurality of discrete point coordinates from the second node. Based on the plurality of discrete point coordinates, the first node may determine a plurality of discrete points from the second data feature, and then calculate the probability density distribution of the second data feature corresponding to the plurality of discrete points. Since the discrete point coordinates are the same, the probability density distribution of the second data feature may be compared with the probability density distribution of the first data feature.

[0224] As an embodiment, when the first node is UE and the second node is LMF, the core of the KDE-based calculation solution is that the LMF end obtains the feature distribution of training data through KDE, and sends the necessary kernel density estimation value, kernel function table index and other information to the UE. The UE completes the calculation of KL divergence and evaluates the difference in feature distribution through KL divergence, thereby monitoring the performance of the first model.

[0225] In the above embodiment, it is necessary to select appropriate kernel functions and bandwidth parameters in the KDE process, and the table index can also be used to send these parameters to reduce the transmission load. Exemplarily, the table index can be predefined in advance, and the LMF sends the kernel function and bandwidth parameters to the UE through the table index. The UE can find the corresponding kernel function parameters according to the index to reuse, without transmitting the complete function and parameters every time.

[0226] As an implementation, the first node may receive a table index for determining a kernel function and a bandwidth parameter from the second node. The kernel function and the bandwidth parameter may be used to determine a probability density function for calculating a probability density distribution. Based on the same kernel function and bandwidth parameter, the first node and the second node use the same algorithm rules to calculate the probability density distribution of the training data and the real-time data, respectively, for comparison.

[0227] In the above embodiment, the LMF side can be responsible for storing the data set used to train the first model, which is the basis for KDE calculation. Through KDE, LMF can calculate the probability density function of the distribution of training set features The calculation of the probability density function involves choosing appropriate kernel functions and bandwidth parameters.

[0228] As an embodiment, the first data feature determined by the training data can be expressed as {x1, x2, ..., x n The first data feature can be extracted from the first model on the second node or the device where the second node is located, and stored on the second node or the device where the second node is located. n represents the number of samples of the feature. The second data feature of the wireless signal data collected in the current period can be expressed as {y1, y2, ..., y m}, extracted by the first model at the first node end and stored at the first node end, and m represents the number of samples of the feature.

[0229] On the second node, use the training data set {x1,x2,...,x n} to estimate the distribution of the feature, that is, the probability density distribution of the first data feature. For each sample point of the first data feature, the Gaussian kernel function (probability density function) can be used to estimate the density distribution:

[0230]

[0231] When the system is running, the real-time features {y1,y2,...,y m} also needs to use KDE to estimate its probability density distribution, that is, the probability density distribution of the second data feature. The first node also uses the kernel function to estimate its density distribution:

[0232]

[0233] As described above, in this embodiment, the difference in feature distribution can be evaluated by KL divergence, that is, the feature distribution difference is KL divergence. As an example, KL divergence can be calculated based on discrete sampling.

[0234] As an example, the coordinates of the plurality of discrete points may be a matrix of size k×d, where d is the dimension of the feature space.

[0235] In this embodiment, the second node can perform discrete point sampling according to the coordinates of multiple discrete points. Multiple discrete points can be selected in the feature space corresponding to the training data. The feature space can be represented by the first data feature. For example, when the second node is LMF, LMF can select a set of discrete points {z1, z2, ..., z k}, where k is the number of discrete points.

[0236] After sampling discrete points, the kernel density estimate can be calculated at each discrete point to obtain the density value set of all discrete points. For example, the kernel density estimate calculated at each discrete point is The set of density values ​​at all discrete points forms a list: This list is the discretized density estimate, also known as the probability density distribution.

[0237] In addition, after the second node determines the discrete point, it can transmit the coordinates of the discrete point and the kernel density estimate (or probability density distribution) to the first node. As an example, LMF needs to transmit the coordinates of the discrete point and the kernel density estimate to the UE. The UE can determine the probability density distribution of the first data feature based on the received kernel density estimate. Among them, the kernel density estimate can be a vector of size k, each value of which is a kernel density estimate corresponding to a discrete point. The kernel density estimate can also be called a probability density value.

[0238] The first node can calculate the kernel density estimate based on the received discrete related data. That is, after the second node receives the discrete point coordinates and the kernel density estimate, it can calculate the kernel density estimate of the discrete point at the position corresponding to the discrete point coordinates. For example, the UE receives multiple discrete point coordinates z sent from the LMF i and the corresponding kernel density estimate The UE collects a feature space of a real-time wireless signal, which can be represented by a second data feature. Discrete points corresponding to the coordinates of these discrete points are selected in the feature space, and kernel density estimates of these discrete points are calculated.

[0239] According to the above calculation results, the first node can calculate the KL divergence according to:

[0240]

[0241] In the second embodiment, the first node can determine the difference between the distribution of training data and real-time data features by calculating the KL divergence. If the KL divergence value exceeds a preset threshold (a first preset threshold or a second preset threshold), it means that there is a significant difference between the distribution of real-time data and the distribution of training data, which may mean that the model performance is biased or abnormal.

[0242] For ease of understanding, the following Fig.10 The process of the second embodiment is schematically described. Fig.10 It is introduced from the perspective of the interaction between the first node and the second node. Fig.10 The monitoring in can be periodic monitoring to optimize the first model.

[0243] In step S1010, the first node receives the probability density distribution of the first data feature sent by the second node. For example, the second node can first obtain the probability density function of the first data feature based on the training data, and then determine multiple kernel density estimation values ​​based on the collected multiple discrete points. (z i ). The list of multiple kernel density estimates is the probability density distribution of the first data feature.

[0244] In step S1020, the first node determines the KL divergence based on the probability density distribution of the first data feature and the probability density distribution of the second data feature. Before determining the KL divergence, the first node may first determine the probability density distribution of the second data feature. Optionally, the first node may receive a table index and / or a plurality of discrete point coordinates from the second node to determine the probability density distribution of the second data feature. For example, the second node sends the kernel density estimate When multiple discrete point coordinates are sent, the first node can determine multiple discrete points from the second data feature according to the multiple discrete point coordinates, and then determine multiple kernel density estimation values ​​corresponding to the second data feature based on these discrete points. Thus, the probability density distribution of the second data feature is obtained. The first node can determine the KL divergence according to the two probability density distributions.

[0245] In the second embodiment, the additional signaling overhead is mainly the coordinates of the discrete points and the corresponding The vector needs to be sent to the first node via signaling. In addition, when the first model needs to be updated, the LMF will re-send the new first model to the UE, including new kernel density estimation parameters. If the first model does not need to be updated, the kernel density estimation value The and / or table index only needs to be sent once and will not be resent until the next model update, thereby reducing signaling overhead.

[0246] Depend on Fig. 9 and Fig.10 It can be seen that the system can continuously collect input data and extract data features, and then periodically use statistical methods such as the MMD algorithm or the KDE algorithm to compare the differences between these feature distributions and the feature distributions during training. If the difference exceeds the preset threshold, the system will determine that there may be problems with the performance of the first model and trigger the adjustment or update mechanism of the first model, thereby ensuring that the first model maintains stable and efficient performance in a dynamic environment.

[0247] The above describes two embodiments in which the feature distribution difference is the data distribution error and the KL divergence. In other embodiments, the feature distribution difference between the first data feature and the second data feature may include the data distribution error and the KL divergence. For example, when the computing power of the first node is relatively strong, the performance of the first model may be monitored based on the two-dimensional data composed of the data distribution error and the KL divergence.

[0248] As an implementation method, the first node may determine the feature distribution difference between the first data feature and the second data feature in different ways in different time periods. For example, the data distribution error of the first data feature and the second data feature is determined in the first time period, and the KL divergence of the first data feature and the second data feature is determined in the second time period. The first time period and the second time period may be different time periods within a monitoring cycle. The first node may perform performance monitoring based on the data distribution error and KL divergence determined in different time periods.

[0249] As another implementation, the first node may determine the feature distribution difference between the first data feature and the second data feature in two different ways in the same period. That is, the first node may calculate the data distribution error of the first data feature and the second data feature and the KL divergence of the first data feature and the second data feature in the same period, and then monitor the first model according to the two parameters.

[0250] In other embodiments, the first node may also determine the feature distribution difference between the first data feature and the second data feature according to one or more other methods to perform performance monitoring.

[0251] Combination of the above Figures 7 to 10 The following describes a method for real-time performance monitoring of the first model based on MMD and KDE. Fig.11 and Fig.12 , describes the embodiments of the present application in more detail. It should be noted that Figures 7 to 10 The examples are only intended to help those skilled in the art understand the embodiments of the present application, and are not intended to limit the embodiments of the present application to the specific values ​​or specific scenarios illustrated. Figures 7 to 10 It is obvious that various equivalent modifications or changes can be made, and such modifications or changes also fall within the scope of the embodiments of the present application.

[0252] Fig.11 This is the signaling process of the MMD-based feature difference calculation scheme. The MMD-based feature difference calculation scheme is used to compare the feature distribution differences between the inference phase and the training phase, so as to evaluate the performance of the first model. Fig.11 It is introduced from the perspective of the interaction between the three entities of gNB, UE and LMF, where UE is the first node and LMF is the second node.

[0253] See also Fig.11 In step S1102, LMF stores the training data. LMF can perform training of the first model according to the training data.

[0254] In step S1104, the LMF executes the distribution of the first model. When the first model needs to be updated, the updated first model is distributed.

[0255] In step S1106, the UE receives RRC signaling sent by the gNB to collect wireless signal characteristics as input data for the first model.

[0256] In step S1108, LMF obtains feature X according to the training data. Feature X is the first data feature.

[0257] In step S1110, LMF calculates and obtains formula (2) in formula ①.

[0258] In step S1112, the UE obtains feature Y according to the wireless signal data collected in real time. Feature Y is the second data feature.

[0259] In step S1114, the UE feeds back RRC signaling to the gNB.

[0260] In step S1116, the UE uploads feature Y to the LMF.

[0261] In step S1118, LMF calculates and obtains equation (1) and equation (3) in formula ①.

[0262] In step S1120, LMF performs feature difference calculation to obtain the MMD result, i.e., the data distribution error.

[0263] In step S1122, the LMF sends the MMD result to the UE.

[0264] In step S1124, the UE obtains the MMD result.

[0265] In step S1126, the UE compares a preset error threshold with the MMD result. Based on the comparison between the threshold and the MMD result, the UE determines whether the first model needs to be updated or modified. The preset error threshold may be the first preset threshold or the second preset threshold described above.

[0266] Fig.12 The signaling process of the KDE-based feature distribution difference calculation scheme. Fig.11 same, Fig.12 It is also introduced from the perspective of the interaction between the three entities of gNB, UE and LMF.

[0267] See also Fig.12, steps S1202 to S1206, S1214 and Figure 6 Steps S1102 to S1106 and S1114 are the same and will not be repeated here.

[0268] In step S1208, LMF uses KDE to estimate the characteristic distribution of the training set data and obtains the probability density function

[0269] In step S1210, LMF selects a set of discrete points {z1, z2, ..., z k}, where k is the number of discrete points, and k discrete points z are obtained i .

[0270] In step S1212, LMF calculates the discrete points z for each discrete point. i The kernel density estimate of The kernel density estimates at these discrete points are collected into a list The list is the probability density distribution of the first data feature.

[0271] In step S1216, LMF sends these discrete points z to the UE. i The coordinates and kernel density estimates of

[0272] In step S1218, the LMF sends a table index including a kernel function and a bandwidth parameter to the UE via RRC signaling.

[0273] In step S1220, the UE obtains a kernel density estimation formula by looking up a table.

[0274] In step S1222, the UE uses these discrete points z i The coordinates of the wireless channel are determined from the characteristics of the wireless channel data (also called the feature space), and the kernel density estimation values ​​of these discrete points are calculated according to the kernel density estimation formula. The kernel density estimates at these discrete points are collected into a list The list is the probability density distribution of the second data feature. The wireless channel data feature is the second data feature based on the wireless signal data of the current time period collected in real time by the UE.

[0275] In step S1224, the UE calculates the KL divergence based on the probability density distribution of the first data feature and the probability density distribution of the second data feature, that is, calculates the KL divergence between the training set data and the wireless signal collected in real time by the UE. If the KL divergence exceeds the preset threshold, the model performance is abnormal and needs to be retrained or adjusted. The preset threshold can be the first preset threshold or the second preset threshold described above.

[0276] Fig.11 and Fig.12 The feature distribution difference calculation methods based on MMD and KDE are introduced respectively. By comparing the two calculation methods, we can see that the advantages and disadvantages of the two methods and their applicable scenarios are shown in Table 1.

[0277] Table 1

[0278]

[0279] In summary, the embodiment of the present application evaluates the performance of the first model by comparing the difference between the input features extracted by the first model in the actual deployment and the feature distribution in the model training stage. Specifically, the system can use techniques such as the maximum mean difference (MMD) or kernel density estimation (KDE) algorithm to calculate the distribution difference between the feature data collected from the deployment environment and the training data. Once these differences exceed a preset threshold, the system will automatically determine that the performance of the first model may be abnormal, and then start the update or adjustment mechanism of the first model.

[0280] When the first model is used to locate the first node, the model monitoring method proposed in the embodiment of the present application significantly reduces the dependence on real location data, and provides a more sensitive and effective performance monitoring method that can quickly adapt to environmental changes to ensure that the first model maintains stability and efficiency in complex and dynamic scenarios.

[0281] The following is an example of the feature difference calculation scheme based on MMD, taking scenario 1 in the previous positioning method as an example. In this scenario, the feature difference calculation scheme based on MMD is used in the operation phase after the model is deployed, that is, the inference phase. By comparing the real-time wireless signal characteristics in the deployment environment with the data characteristics in the training phase, it is possible to evaluate whether the positioning performance of the first model deviates.

[0282] First, in the system initialization phase, the first model can be trained on the second node. During the training process, the first data feature X can be extracted. After the training is completed, it is sent to the first node end through signaling for deployment.

[0283] Secondly, the first node can perform real-time feature collection based on the wireless signal data in the current environment to obtain the second data feature Y.

[0284] Again, the first node can transmit the data feature. In order to evaluate the difference between the deployment environment and the training environment, the first node uploads the second data feature Y currently extracted to the first node through signaling. The uploaded features can be performed periodically, and the frequency can be adjusted according to the system settings and the complexity of the communication environment.

[0285] Again, the first node performs relevant calculations based on the MMD algorithm to evaluate the performance of the first model. The first node can calculate the distribution difference between the second data feature Y collected in real time and the first data feature X of the training data according to formula ①.

[0286] Finally, the first node returns the result to the second node and performs the model adaptation mechanism. After the first node completes the calculation, the MMD result is fed back to the second node through signaling. The second node compares the data distribution error representing the MMD result with the preset threshold. If the data distribution error is less than the preset threshold, it means that the feature distribution of the data collected from the real-time environment is roughly consistent with the data in the training phase, and the positioning performance of the first model is normal. If the data distribution error exceeds the preset threshold, it means that the environmental change has caused the model performance to deteriorate. At this time, the system can trigger a model update or other correction mechanism, such as parameter fine-tuning or retraining.

[0287] During the implementation of the above method, periodic monitoring and system optimization of the first model can be achieved. When the model performance remains normal, the system only needs to periodically collect wireless signal data, extract features, and calculate the data distribution error between the first data feature and the second data feature through the MMD algorithm, without frequently updating the model, thereby effectively reducing signaling overhead and computing burden. When significant environmental changes are detected, the system will perform necessary model adjustments and optimizations to ensure positioning accuracy and system stability.

[0288] Through this embodiment, the system can effectively evaluate the performance of the first model deployed at the first node end under scenario 1, and monitor the impact of environmental changes on model performance in real time through MMD calculation, ensuring that the location information output by the first node is always highly accurate and robust.

[0289] Combination of the above Figures 1 to 12 , describes the method embodiment of the present application in detail, and the following is combined with Figures 13 to 16 , describes the device embodiment of the present application in detail. It should be understood that the description of the method embodiment corresponds to the description of the device embodiment, so the parts not described in detail can refer to the previous method embodiment.

[0290] Fig.13 A first node for wireless communication provided in an embodiment of the present application. The first node may be a terminal device or a network device. Fig.13 As shown, the first node 1300 includes a first processor 1310 .

[0291] The first processor 1310 can be used to monitor the performance of the first model based on the difference in feature distribution between the first data feature and the second data feature; wherein the first data feature is an intermediate feature of the first model in the training stage, and the second data feature is an intermediate feature of the first model in the reasoning stage.

[0292] As an embodiment, the first model includes K intermediate layers, K is a positive integer, the first data feature and the second data feature are both intermediate features output by the first intermediate layer of the first model, and the first intermediate layer is one of the K intermediate layers.

[0293] As an embodiment, the first intermediate layer is directly connected to the output layer of the first model.

[0294] As an embodiment, the first processor 1310 is also used to trigger an update mechanism of the first model when the feature distribution difference is greater than or equal to a first preset threshold; or, when the feature distribution difference is greater than or equal to a second preset threshold, trigger a correction mechanism of the first model.

[0295] As an embodiment, the feature distribution difference includes at least one of the following: a data distribution error between the first data feature and the second data feature determined by a maximum mean difference algorithm; a KL divergence between the first data feature and the second data feature determined by a kernel density estimation algorithm.

[0296] As an embodiment, the characteristic distribution difference is the data distribution error, and the first processor 1310 is also used to determine the second data characteristic; the first node 1300 also includes a first transceiver, which can be used to send the second data characteristic to the second node, and the second data characteristic is used by the second node to determine the data distribution error; the first transceiver is also used to receive the data distribution error from the second node.

[0297] As an embodiment, the feature distribution difference is the KL divergence, and the first node 1300 also includes a second transceiver, which can be used to receive the probability density distribution of the first data feature from the second node; the first processor 1310 is also used to determine the probability density distribution of the second data feature; and the KL divergence is determined based on the probability density distribution of the first data feature and the probability density distribution of the second data feature.

[0298] As an embodiment, the second transceiver is also used to receive a table index from the second node for determining a kernel function and a bandwidth parameter, wherein the kernel function and the bandwidth parameter are used to determine a probability density distribution of the first data feature and a probability density distribution of the second data feature.

[0299] As an embodiment, the second transceiver is also used to receive multiple discrete point coordinates from the second node; the first processor 1310 is also used to: determine multiple discrete points from the second data feature based on the multiple discrete point coordinates; and calculate the probability density distribution of the second data feature corresponding to the multiple discrete points.

[0300] As an embodiment, the first processor 1310 is further used to determine the wireless signal data of the current time period; and input the wireless signal data into the first model to determine the second data feature corresponding to the wireless signal data of the current time period.

[0301] As an embodiment, the first node 1300 further includes a third transceiver, which can be used to receive the first model from the second node; wherein the first model is used to locate the first node.

[0302] As an embodiment, the first model is an AI / ML model.

[0303] As an embodiment, the first processor 1310 may be a processor 1510. The first node 1300 may further include a memory 1520 and a transceiver 1530, as shown in FIG. Fig.15 shown.

[0304] Fig.14 A second node for wireless communication provided in an embodiment of the present application. The second node may be a device or entity used for positioning on the network side, such as LMF. Fig.14 As shown, the second node 1400 includes a second processor 1410 .

[0305] The second processor 1410 can be used to train the first model; wherein the feature distribution difference between the first data feature and the second data feature is used to monitor the performance of the first model; the first data feature is an intermediate feature of the first model in the training stage, and the second data feature is an intermediate feature of the first model in the reasoning stage.

[0306] As an embodiment, the first model includes K intermediate layers, K is a positive integer, the first data feature and the second data feature are both intermediate features output by the first intermediate layer of the first model, and the first intermediate layer is one of the K intermediate layers.

[0307] As an embodiment, the first intermediate layer is directly connected to the output layer of the first model.

[0308] As an embodiment, the second processor 1410 is also used to receive an update request for the first model triggered by the first node when the feature distribution difference is greater than or equal to a first preset threshold; or, when the feature distribution difference is greater than or equal to a second preset threshold, receive a correction request for the first model triggered by the first node.

[0309] As an embodiment, the feature distribution difference includes at least one of the following: a data distribution error between the first data feature and the second data feature determined by a maximum mean difference algorithm; a KL divergence between the first data feature and the second data feature determined by a kernel density estimation algorithm.

[0310] As an embodiment, the characteristic distribution difference is the data distribution error, and the second node 1400 also includes a fourth transceiver, which can be used to receive the second data characteristic from the first node; the second processor 1410 is also used to determine the data distribution error based on the first data characteristic and the second data characteristic; the fourth transceiver is also used to send the data distribution error to the first node.

[0311] As an embodiment, the feature distribution difference is the KL divergence, and the second processor 1410 is also used to determine the probability density distribution of the first data feature; the second node 1400 also includes a fifth transceiver, which can be used to send the probability density distribution of the first data feature to the first node.

[0312] As an embodiment, the fifth transceiver is also used to send a table index for determining a kernel function and a bandwidth parameter to the first node, wherein the kernel function and the bandwidth parameter are used to determine a probability density distribution of the first data feature and a probability density distribution of the second data feature.

[0313] As an embodiment, the fifth transceiver is also used to send multiple discrete point coordinates to the first node, and the multiple discrete point coordinates are used by the first node to determine multiple discrete points from the second data feature, and the multiple discrete points are used to calculate the probability density distribution of the second data feature.

[0314] As an embodiment, the second processor 1410 is further configured to input training data into the first model to determine a first data feature.

[0315] As an embodiment, the second node 1400 further includes a sixth transceiver, which can be used to send the first model to the first node; wherein the first model is used to locate the first node.

[0316] As an embodiment, the first model is an AI / ML model.

[0317] As an embodiment, the second processor 1410 may be the processor 1510. The second node 1400 may further include a memory 1520 and a transceiver 1530. Fig.15 shown.

[0318] Fig.15 It is a schematic structural diagram of a communication device according to an embodiment of the present application. Fig.15 The dotted line in the figure indicates that the unit or module is optional. The device 1500 can be used to implement the method described in the above method embodiment. The device 1500 can be a chip, a user equipment or a network device.

[0319] The device 1500 may include one or more processors 1510. The processor 1510 may support the device 1500 to implement the method described in the method embodiment above. The processor 1510 may be a general-purpose processor or a special-purpose processor. For example, the processor may be a central processing unit (CPU). Alternatively, the processor may also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0320] The apparatus 1500 may further include one or more memories 1520. The memory 1520 stores a program, which can be executed by the processor 1510, so that the processor 1510 executes the method described in the above method embodiment. The memory 1520 may be independent of the processor 1510 or integrated in the processor 1510.

[0321] The apparatus 1500 may further include a transceiver 1530. The processor 1510 may communicate with other devices or chips through the transceiver 1530. For example, the processor 1510 may transmit and receive data with other devices or chips through the transceiver 1530.

[0322] Fig.16 The hardware module diagram of the communication device provided in the embodiment of the present application. Specifically, Fig.16 A block diagram is shown of a first communication device 1650 and a second communication device 1610 communicating with each other in an access network.

[0323] The first communication device 1650 includes a controller / processor 1659, a memory 1660, a data source 1667, a transmit processor 1668, a receive processor 1656, a multi-antenna transmit processor 1657, a multi-antenna receive processor 1658, a transmitter / receiver 1654 and an antenna 1652.

[0324] The second communication device 1610 includes a controller / processor 1675, a memory 1676, a data source 1677, a receive processor 1670, a transmit processor 1616, a multi-antenna receive processor 1672, a multi-antenna transmit processor 1671, a transmitter / receiver 1618 and an antenna 1620.

[0325] In the transmission from the second communication device 1610 to the first communication device 1650, at the second communication device 1610, the upper layer data packets from the core network or the upper layer data packets from the data source 1677 are provided to the controller / processor 1675. The core network and the data source 1677 represent all the protocol layers above the L2 layer. The controller / processor 1675 implements the functionality of the L2 layer. In the transmission from the second communication device 1610 to the first communication device 1650, the controller / processor 1675 provides header compression, encryption, packet segmentation and reordering, multiplexing between logical and transport channels, and radio resource allocation to the first communication device 1650 based on various priority metrics. The controller / processor 1675 is also responsible for the retransmission of lost packets and signaling to the first communication device 1650. The transmit processor 1616 and the multi-antenna transmit processor 1671 implement various signal processing functions for the L1 layer (i.e., the physical layer). The transmit processor 1616 implements coding and interleaving to facilitate forward error correction at the second communication device 1610, as well as mapping of signal constellations based on various modulation schemes (e.g., binary phase shift keying, quadrature phase shift keying, M phase shift keying, M quadrature amplitude modulation). The multi-antenna transmit processor 1671 performs digital spatial precoding, including codebook-based precoding and non-codebook-based precoding, and beamforming processing on the coded and modulated symbols to generate one or more spatial streams. The transmit processor 1616 then maps each spatial stream to a subcarrier, multiplexes with a reference signal (e.g., a pilot) in the time domain and / or frequency domain, and then uses an inverse fast Fourier transform to generate a physical channel carrying a time domain multi-carrier symbol stream. The multi-antenna transmit processor 1671 then performs a transmit analog precoding / beamforming operation on the time domain multi-carrier symbol stream. Each transmitter 1618 converts the baseband multi-carrier symbol stream provided by the multi-antenna transmit processor 1671 into a radio frequency stream, and then provides it to a different antenna 1620.

[0326] In the transmission from the second communication device 1610 to the first communication device 1650, at the first communication device 1650, each receiver 1654 receives a signal through its corresponding antenna 1652. Each receiver 1654 recovers the information modulated onto the radio frequency carrier and converts the radio frequency stream into a baseband multi-carrier symbol stream and provides it to the receiving processor 1656. The receiving processor 1656 and the multi-antenna receiving processor 1658 implement various signal processing functions of the L1 layer. The multi-antenna receiving processor 1658 performs a receiving analog precoding / beamforming operation on the baseband multi-carrier symbol stream from the receiver 1654. The receiving processor 1656 uses a fast Fourier transform to convert the baseband multi-carrier symbol stream after the receiving analog precoding / beamforming operation from the time domain to the frequency domain. In the frequency domain, the physical layer data signal and the reference signal are demultiplexed by the receiving processor 1656, wherein the reference signal will be used for channel estimation, and the data signal is recovered after multi-antenna detection in the multi-antenna receiving processor 1658 to any spatial stream with the first communication device 1650 as the destination. The symbols on each spatial stream are demodulated and recovered in the receiving processor 1656, and soft decisions are generated. The receiving processor 1656 then decodes and deinterleaves the soft decisions to recover the upper layer data and control signals transmitted by the second communication device 1610 on the physical channel. The upper layer data and control signals are then provided to the controller / processor 1659. The controller / processor 1659 implements the functions of the L2 layer. The controller / processor 1659 may be associated with a memory 1660 that stores program codes and data. The memory 1660 may be referred to as a computer-readable medium. In the transmission from the second communication device 1610 to the first communication device 1650, the controller / processor 1659 provides multiplexing between transport and logical channels, packet reassembly, decryption, header decompression, control signal processing to recover the upper layer data packets from the second communication device 1610. The upper layer data packets are then provided to all protocol layers above the L2 layer. Various control signals may also be provided to the L3 for L3 processing.

[0327] In the transmission from the first communication device 1650 to the second communication device 1610, at the first communication device 1650, the upper layer data packets are provided to the controller / processor 1659 using the data source 1667. The data source 1667 represents all the protocol layers above the L2 layer. Similar to the transmission function at the second communication device 1610 described in the transmission from the second communication device 1610 to the first communication device 1650, the controller / processor 1659 implements header compression, encryption, packet segmentation and reordering, and multiplexing between logical and transport channels, and implements L2 layer functions for the user plane and the control plane. The controller / processor 1659 is also responsible for the retransmission of lost packets and signaling to the second communication device 1610. The transmit processor 1668 performs modulation mapping and channel coding processing, and the multi-antenna transmit processor 1657 performs digital multi-antenna spatial precoding, including codebook-based precoding and non-codebook-based precoding, and beamforming processing. Then, the transmit processor 1668 modulates the generated spatial stream into a multi-carrier / single-carrier symbol stream, which is then provided to different antennas 1652 via the transmitter 1654 after analog precoding / beamforming operations in the multi-antenna transmit processor 1657. Each transmitter 1654 first converts the baseband symbol stream provided by the multi-antenna transmit processor 1657 into a radio frequency symbol stream, and then provides it to the antenna 1652.

[0328] In the transmission from the first communication device 1650 to the second communication device 1610, the function at the second communication device 1610 is similar to the reception function at the first communication device 1650 described in the transmission from the second communication device 1610 to the first communication device 1650. Each receiver 1618 receives a radio frequency signal through its corresponding antenna 1620, converts the received radio frequency signal into a baseband signal, and provides the baseband signal to the multi-antenna reception processor 1672 and the reception processor 1670. The reception processor 1670 and the multi-antenna reception processor 1672 jointly implement the functions of the L1 layer. The controller / processor 1675 implements the L2 layer functions. The controller / processor 1675 can be associated with a memory 1676 that stores program codes and data. The memory 1676 can be referred to as a computer-readable medium. In transmission from the first communication device 1650 to the second communication device 1610, the controller / processor 1675 provides demultiplexing between transport and logical channels, packet reassembly, decryption, header decompression, and control signal processing to recover upper layer data packets from the first communication device 1650. The upper layer data packets from the controller / processor 1675 can be provided to the core network or all protocol layers above the L2 layer, and various control signals can also be provided to the core network or L3 for L3 processing.

[0329] As an embodiment, the first communication device 1650 apparatus includes: at least one processor and at least one memory, the at least one memory including computer program code; the at least one memory and the computer program code are configured to be used together with the at least one processor, and the first communication device 1650 apparatus at least: monitors the performance of the first model based on the feature distribution difference between the first data feature and the second data feature; wherein the first data feature is an intermediate feature of the first model in the training stage, and the second data feature is an intermediate feature of the first model in the reasoning stage.

[0330] As an embodiment, the first communication device 1650 apparatus includes: a memory storing a computer-readable instruction program, wherein the computer-readable instruction program generates an action when executed by at least one processor, and the action includes: monitoring the performance of the first model based on the feature distribution difference between the first data feature and the second data feature; wherein the first data feature is an intermediate feature of the first model in the training stage, and the second data feature is an intermediate feature of the first model in the reasoning stage.

[0331] As an embodiment, the first communication device 1650 corresponds to the first node in this application.

[0332] As an embodiment, the second communication device 1610 corresponds to the second node in this application.

[0333] As an embodiment, the first communication device 1650 is a user equipment, and the user equipment can serve as a relay node.

[0334] As an embodiment, the first communication device 1650 is a network controlled relay (NCR).

[0335] As an embodiment, the first communication device 1650 is a relay wireless repeater.

[0336] As an embodiment, the first communication device 1650 is a relay.

[0337] As an embodiment, the second communication device 1610 is a location management function (LMF).

[0338] As an embodiment, the first communication device 1650 corresponds to the first node in the present application, and the controller / processor 1659 is used to monitor the performance of the first model according to the difference in feature distribution between the first data feature and the second data feature.

[0339] The embodiment of the present application also provides a computer-readable storage medium for storing a program. The computer-readable storage medium can be applied to a terminal or network device provided in the embodiment of the present application, and the program enables a computer to execute the method performed by the terminal device, network device or core network entity in each embodiment of the present application.

[0340] The embodiment of the present application also provides a computer program product. The computer program product includes a program. The computer program product can be applied to the terminal or network device provided in the embodiment of the present application, and the program enables the computer to execute the method performed by the terminal device, network device or core network entity in each embodiment of the present application.

[0341] The embodiment of the present application also provides a computer program. The computer program can be applied to the terminal or network device provided in the embodiment of the present application, and the computer program enables the computer to execute the method performed by the terminal device, network device or core network entity in each embodiment of the present application.

[0342] It should be understood that the terms "system" and "network" in this application can be used interchangeably. In addition, the terms used in this application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application. The terms "first", "second", "third" and "fourth" in the specification and claims of this application and the accompanying drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions.

[0343] In the embodiments of the present application, the "indication" mentioned can be a direct indication, an indirect indication, or an indication of an association relationship. For example, A indicates B, which can mean that A directly indicates B, for example, B can be obtained through A; it can also mean that A indirectly indicates B, for example, A indicates C, and B can be obtained through C; it can also mean that there is an association relationship between A and B.

[0344] In the embodiment of the present application, "B corresponding to A" means that B is associated with A, and B can be determined according to A. However, it should be understood that determining B according to A does not mean determining B only according to A, and B can also be determined according to A and / or other information.

[0345] In the embodiments of the present application, the term "corresponding" may indicate that there is a direct or indirect correspondence between the two, or an association relationship between the two, or a relationship of indication and being indicated, configuration and being configured, etc.

[0346] In the embodiments of the present application, "pre-definition" or "pre-configuration" can be implemented by pre-saving corresponding codes, tables or other methods that can be used to indicate relevant information in a device (for example, including a user device and a network device), and the present application does not limit the specific implementation method. For example, pre-definition can refer to what is defined in the protocol.

[0347] In the embodiments of the present application, the “protocol” may refer to a standard protocol in the communication field, for example, it may include an LTE protocol, an NR protocol, and related protocols used in future communication systems, and the present application does not limit this.

[0348] In the embodiments of the present application, the term "and / or" is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0349] In various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0350] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0351] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0352] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0353] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that can be read by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a digital video disc (DVD)), or a semiconductor medium (eg, a solid state disk (SSD)).

[0354] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method in a first node for wireless communication, characterized in that include: Monitoring the performance of the first model according to a difference in feature distribution between the first data feature and the second data feature; The first data feature is an intermediate feature of the first model in the training phase, and the second data feature is an intermediate feature of the first model in the reasoning phase.

2. The method according to claim 1, characterized in that The first model includes K intermediate layers, K is a positive integer, the first data feature and the second data feature are both intermediate features output by the first intermediate layer of the first model, and the first intermediate layer is one of the K intermediate layers.

3. The method according to claim 2, characterized in that The first intermediate layer is directly connected to the output layer of the first model.

4. The method according to any one of claims 1 to 3, characterized in that The method further comprises: When the feature distribution difference is greater than or equal to a first preset threshold, triggering an update mechanism of the first model; or, When the feature distribution difference is greater than or equal to a second preset threshold, a correction mechanism of the first model is triggered.

5. The method according to any one of claims 1 to 4, characterized in that The feature distribution difference includes at least one of the following: a data distribution error between the first data feature and the second data feature determined by a maximum mean difference algorithm; The KL divergence of the first data feature and the second data feature is determined by a kernel density estimation algorithm.

6. The method according to claim 5, characterized in that The feature distribution difference is the data distribution error, and the method further includes: determining the second data feature; Sending the second data feature to a second node, where the second data feature is used by the second node to determine the data distribution error; The data distribution error is received from the second node.

7. The method according to claim 5, characterized in that The feature distribution difference is the KL divergence, and the method further includes: receiving a probability density distribution of the first data feature from a second node; determining a probability density distribution of the second data feature; The KL divergence is determined according to the probability density distribution of the first data feature and the probability density distribution of the second data feature.

8. The method according to claim 7, characterized in that The method further comprises: A table index for determining a kernel function and a bandwidth parameter is received from the second node, wherein the kernel function and the bandwidth parameter are used to determine a probability density distribution of the first data feature and a probability density distribution of the second data feature.

9. The method according to claim 7 or 8, characterized in that: The method further comprises: receiving a plurality of discrete point coordinates from the second node; Determining a plurality of discrete points from the second data feature according to the plurality of discrete point coordinates; Calculate the probability density distribution of the second data feature corresponding to the multiple discrete points.

10. The method according to any one of claims 1 to 9, characterized in that The method further comprises: Determine wireless signal data for the current period; The wireless signal data is input into the first model to determine the second data feature corresponding to the wireless signal data in the current time period.

11. The method according to any one of claims 1 to 10, characterized in that The method further comprises: receiving the first model from a second node; The first model is used to locate the first node.

12. The method according to any one of claims 1 to 11, characterized in that The first model is an AI / ML model.

13. A method in a second node for wireless communication, characterized in that: include: Train the first model; Among them, the feature distribution difference between the first data feature and the second data feature is used to monitor the performance of the first model; the first data feature is the intermediate feature of the first model in the training stage, and the second data feature is the intermediate feature of the first model in the reasoning stage.

14. The method according to claim 13, characterized in that The first model includes K intermediate layers, K is a positive integer, the first data feature and the second data feature are both intermediate features output by the first intermediate layer of the first model, and the first intermediate layer is one of the K intermediate layers.

15. The method according to claim 14, characterized in that The first intermediate layer is directly connected to the output layer of the first model.

16. The method according to any one of claims 13 to 15, characterized in that The method further comprises: When the feature distribution difference is greater than or equal to a first preset threshold, receiving an update request for the first model triggered by a first node; or, When the feature distribution difference is greater than or equal to a second preset threshold, a correction request for the first model triggered by the first node is received.

17. The method according to any one of claims 13 to 16, characterized in that The feature distribution difference includes at least one of the following: a data distribution error between the first data feature and the second data feature determined by a maximum mean difference algorithm; The KL divergence of the first data feature and the second data feature is determined by a kernel density estimation algorithm.

18. The method according to claim 17, characterized in that The feature distribution difference is the data distribution error, and the method further includes: receiving the second data feature from the first node; determining the data distribution error according to the first data feature and the second data feature; The data distribution error is sent to the first node.

19. The method according to claim 17, characterized in that The feature distribution difference is the KL divergence, and the method further includes: determining a probability density distribution of the first data feature; The probability density distribution of the first data feature is sent to the first node.

20. The method according to claim 19, characterized in that The method further comprises: A table index for determining a kernel function and a bandwidth parameter is sent to the first node, where the kernel function and the bandwidth parameter are used to determine a probability density distribution of the first data feature and a probability density distribution of the second data feature.

21. The method according to claim 19 or 20, characterized in that The method further comprises: A plurality of discrete point coordinates are sent to the first node, where the plurality of discrete point coordinates are used by the first node to determine a plurality of discrete points from the second data feature, where the plurality of discrete points are used to calculate the probability density distribution of the second data feature.

22. The method according to any one of claims 13 to 21, characterized in that The method further comprises: Training data is input into the first model to determine the first data feature.

23. The method according to any one of claims 13 to 22, characterized in that The method further comprises: Sending the first model to a first node; The first model is used to locate the first node.

24. The method according to any one of claims 13 to 23, characterized in that The first model is an AI / ML model.

25. A first node for wireless communication, characterized in that: include: A first processor is used to monitor the performance of the first model according to a feature distribution difference between the first data feature and the second data feature; The first data feature is an intermediate feature of the first model in the training phase, and the second data feature is an intermediate feature of the first model in the reasoning phase.

26. The first node according to claim 25, characterized in that The first model includes K intermediate layers, K is a positive integer, the first data feature and the second data feature are both intermediate features output by the first intermediate layer of the first model, and the first intermediate layer is one of the K intermediate layers.

27. The first node according to claim 26, characterized in that The first intermediate layer is directly connected to the output layer of the first model.

28. The first node according to any one of claims 25-27, characterized in that: The first processor is further configured to: When the feature distribution difference is greater than or equal to a first preset threshold, triggering an update mechanism of the first model; or, When the feature distribution difference is greater than or equal to a second preset threshold, a correction mechanism of the first model is triggered.

29. The first node according to any one of claims 25-28, characterized in that: The feature distribution difference includes at least one of the following: a data distribution error between the first data feature and the second data feature determined by a maximum mean difference algorithm; The KL divergence of the first data feature and the second data feature is determined by a kernel density estimation algorithm.

30. The first node according to claim 29, characterized in that The feature distribution difference is the data distribution error, and the first processor is further used to determine the second data feature; The first node also includes: A first transceiver, configured to send the second data feature to a second node, wherein the second data feature is used by the second node to determine the data distribution error; The first transceiver is further configured to receive the data distribution error from the second node.

31. The first node according to claim 29, characterized in that The feature distribution difference is the KL divergence, and the first node further includes: A second transceiver, configured to receive the probability density distribution of the first data feature from a second node; The first processor is further configured to determine a probability density distribution of the second data feature; and determine the KL divergence according to the probability density distribution of the first data feature and the probability density distribution of the second data feature.

32. The first node according to claim 31, characterized in that The second transceiver is also used to receive a table index for determining a kernel function and a bandwidth parameter from the second node, wherein the kernel function and the bandwidth parameter are used to determine a probability density distribution of the first data feature and a probability density distribution of the second data feature.

33. The first node according to claim 31 or 32, characterized in that: The second transceiver is also used to receive a plurality of discrete point coordinates from the second node; The first processor is further configured to: Determining a plurality of discrete points from the second data feature according to the plurality of discrete point coordinates; Calculate the probability density distribution of the second data feature corresponding to the multiple discrete points.

34. The first node according to any one of claims 25-33, characterized in that: The first processor is further configured to: Determine wireless signal data for the current period; The wireless signal data is input into the first model to determine the second data feature corresponding to the wireless signal data in the current time period.

35. The first node according to any one of claims 25-34, characterized in that: The first node also includes: a third transceiver, configured to receive the first model from a second node; The first model is used to locate the first node.

36. The first node according to any one of claims 25-35, characterized in that: The first model is an AI / ML model.

37. A second node for wireless communication, characterized in that: include: a second processor, configured to train the first model; Among them, the feature distribution difference between the first data feature and the second data feature is used to monitor the performance of the first model; the first data feature is the intermediate feature of the first model in the training stage, and the second data feature is the intermediate feature of the first model in the reasoning stage.

38. The second node according to claim 37, characterized in that: The first model includes K intermediate layers, K is a positive integer, the first data feature and the second data feature are both intermediate features output by the first intermediate layer of the first model, and the first intermediate layer is one of the K intermediate layers.

39. The second node according to claim 38, characterized in that: The first intermediate layer is directly connected to the output layer of the first model.

40. The second node according to any one of claims 37-39, characterized in that: The second processor is further configured to: When the feature distribution difference is greater than or equal to a first preset threshold, receiving an update request for the first model triggered by a first node; or, When the feature distribution difference is greater than or equal to a second preset threshold, a correction request for the first model triggered by the first node is received.

41. The second node according to any one of claims 37-40, characterized in that: The feature distribution difference includes at least one of the following: a data distribution error between the first data feature and the second data feature determined by a maximum mean difference algorithm; The KL divergence of the first data feature and the second data feature is determined by a kernel density estimation algorithm.

42. The second node according to claim 41, characterized in that The feature distribution difference is the data distribution error, and the second node further includes: a fourth transceiver, configured to receive the second data feature from the first node; The second processor is further configured to determine the data distribution error according to the first data feature and the second data feature; The fourth transceiver is further configured to send the data distribution error to the first node.

43. The second node according to claim 41, characterized in that The feature distribution difference is the KL divergence, and the second processor is further used to determine the probability density distribution of the first data feature; The second node also includes: A fifth transceiver is used to send the probability density distribution of the first data feature to the first node.

44. The second node according to claim 43, characterized in that: The fifth transceiver is further used to send a table index for determining a kernel function and a bandwidth parameter to the first node, where the kernel function and the bandwidth parameter are used to determine a probability density distribution of the first data feature and a probability density distribution of the second data feature.

45. The second node according to claim 43 or 44, characterized in that: The fifth transceiver is also used to send a plurality of discrete point coordinates to the first node, the plurality of discrete point coordinates are used by the first node to determine a plurality of discrete points from the second data feature, and the plurality of discrete points are used to calculate the probability density distribution of the second data feature.

46. ​​The second node according to any one of claims 37-45, characterized in that: The second processor is further configured to input training data into the first model to determine the first data feature.

47. The second node according to any one of claims 37-46, characterized in that: The second node also includes: a sixth transceiver, configured to send the first model to the first node; The first model is used to locate the first node.

48. The second node according to any one of claims 37-47, characterized in that: The first model is an AI / ML model.

49. A node used for wireless communication, characterized in that: It includes a transceiver, a memory and a processor, wherein the memory is used to store programs, and the processor is used to call the programs in the memory and control the transceiver to receive or send signals so that the node executes the method as described in any one of claims 1-12 or 13-24.

50. A device, characterized in that It comprises a processor, which is used to call a program from a memory so that the device executes the method as described in any one of claims 1-12 or 13-24.

51. A chip, characterized in that: It comprises a processor, which is used to call a program from a memory, so that a device equipped with the chip executes the method as described in any one of claims 1-12 or 13-24.

52. A computer-readable storage medium, characterized in that: A program is stored thereon, the program causing a computer to execute the method according to any one of claims 1-12 or 13-24.

53. A computer program product, characterized in that The method comprises a program which causes a computer to execute the method as claimed in any one of claims 1 to 12 or 13 to 24.

54. A computer program, characterized in that The computer program causes a computer to execute the method according to any one of claims 1-12 or 13-24.

Images