Radio environment perception method and apparatus, electronic device, and storage medium

Abstract

Embodiments of the present application provide a radio environment perception method and device, electronic equipment and storage medium, belonging to the field of communication. The method comprises: tokenizing a wireless measurement report to obtain a first token and a second token; inputting the first token and the second token into a radio environment perception model, the model comprising a backbone network, a prediction head and a projection head, the backbone network comprising a hierarchical attention network and a hybrid expert network; performing hierarchical attention calculation on the first token and the second token through the hierarchical attention network to obtain a cell-level mask attention feature; extracting a channel feature of the radio environment based on the cell-level mask attention feature through the hybrid expert network; predicting a reference signal received power of a mask beam based on the channel feature through the prediction head, and extracting a projection embedding feature of the mask beam based on the channel feature through the projection head, which can reduce the difficulty of radio environment perception.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a radio environment sensing method, apparatus, electronic device, and storage medium. Background Technology

[0002] In related technologies, radio environment awareness is achieved based on Channel State Information (CSI) through wireless fundamental models, including channel estimation, CSI feedback, and multi-user precoding. However, collecting high-quality CSI data requires proprietary hardware and complex data processing procedures. With the densification of wireless networks and the trend towards large-scale multiple-input multiple-output (MIMO), the channel dimension is growing exponentially, making real-time acquisition of CSI data for all wireless links virtually impossible in practical engineering. This significantly increases the difficulty of radio environment awareness. Summary of the Invention

[0003] The main objective of this application is to provide a radio environment sensing method, apparatus, electronic device, and storage medium, which aims to reduce the difficulty of radio environment sensing.

[0004] To achieve the above objectives, a first aspect of this application proposes a radio environment sensing method, the method comprising:

[0005] Obtain a wireless measurement report of a radio environment; wherein the radio environment includes at least one cell having at least two beams, the beams including an unmasked beam and a masked beam, and the wireless measurement report includes a first measurement report of the unmasked beam and a second measurement report of the masked beam;

[0006] The wireless measurement report is tokenized to obtain a first token for the first measurement report and a second token for the second measurement report;

[0007] The first token and the second token are input into the radio environment awareness model; wherein, the radio environment awareness model includes a backbone network, a prediction head and a projection head, the prediction head and the projection head are respectively connected to the backbone network, and the backbone network includes a hierarchical attention network and a hybrid expert network;

[0008] The first token and the second token are subjected to hierarchical attention calculation through the hierarchical attention network to obtain cell-level mask attention features;

[0009] The hybrid expert network extracts channel features of the radio environment based on the cell-level mask attention features;

[0010] The prediction head predicts the reference signal received power of the masked beam based on the channel characteristics, and the projection head extracts the projection embedding features of the masked beam based on the channel characteristics.

[0011] In some embodiments, the radio environment awareness model is trained according to the following steps:

[0012] Obtain a sample wireless measurement report; wherein, the sample wireless measurement report includes a first sample measurement report of an unmasked sample beam and a second sample measurement report of a masked sample beam, the second sample measurement report including the sample reference signal received power and physical configuration information of the masked sample beam;

[0013] Wireless feature extraction is performed on the first sample measurement report and the physical configuration information using a preset model to obtain global mask features, the predicted reference signal received power of the mask sample beam, and the predicted projection embedding features.

[0014] The target loss value is obtained by calculating the loss based on the received power of the sample reference signal, the global mask features, the received power of the predicted reference signal, and the predicted projection embedding features.

[0015] The preset model is updated based on the target loss value to obtain the radio environment perception model.

[0016] In some embodiments, the preset model includes a student encoder, a teacher encoder, a preset prediction head, a first projection head, and a second projection head. The student encoder is connected to both the preset prediction head and the first projection head, and the teacher encoder is connected to the second projection head. The step of performing wireless feature extraction on the first sample measurement report and the physical configuration information using the preset model to obtain global mask features, the predicted reference signal received power of the mask sample beam, and the predicted projection embedding features includes:

[0017] The student encoder performs feature encoding on the first sample measurement report and the physical configuration information to obtain encoded features; wherein, the encoded features include the global mask feature and a first feature, the first feature being obtained based on the global mask feature;

[0018] The preset prediction head outputs the received power of the prediction reference signal based on the first feature;

[0019] The first projection head outputs a first projection embedding feature based on the first feature;

[0020] The first sample measurement report and the physical configuration information are enhanced to obtain an enhanced view;

[0021] The enhanced view is feature-encoded using the teacher encoder to obtain the second feature;

[0022] The second projection head outputs a second projection embedding feature based on the second feature;

[0023] The predicted projection embedding feature is determined based on the first projection embedding feature and the second projection embedding feature.

[0024] In some embodiments, the step of calculating the target loss value based on the received power of the sample reference signal, the global mask feature, the received power of the predicted reference signal, and the predicted projection embedding feature includes:

[0025] A first loss value is calculated based on the received power of the sample reference signal, the received power of the predicted reference signal, and the predicted projection embedding features;

[0026] Calculate the second loss value based on the predicted reference signal received power;

[0027] Based on the global mask features, the strongest beam prediction is performed to obtain the expert routing probability;

[0028] Calculate the third loss value based on the expert routing probability;

[0029] The target loss value is obtained by weighting the first loss value, the second loss value, and the third loss value.

[0030] In some embodiments, the predicted projection embedding feature includes a first projection embedding feature and a second projection embedding feature, and the step of calculating a first loss value based on the sample reference signal received power, the predicted reference signal received power, and the predicted projection embedding feature includes:

[0031] The reconstruction loss is calculated based on the received power of the sample reference signal and the received power of the predicted reference signal;

[0032] The distillation loss is calculated based on the first projection embedding feature and the second projection embedding feature;

[0033] The first loss value is calculated based on the reconstruction loss and the distillation loss.

[0034] In some embodiments, the hierarchical attention network includes a first network layer, an average pooling layer, and a second network layer. The step of performing hierarchical attention calculations on the first token and the second token through the hierarchical attention network to obtain cell-level mask attention features includes:

[0035] The first attention calculation is performed on the first token and the second token by the first network layer to obtain the beam-level mask attention features in the cell.

[0036] The average pooling layer is used to perform cell-level beam aggregation on the beam-level mask attention features within the cell to obtain cell-level embedding features.

[0037] The cell-level mask attention feature is obtained by performing a second attention calculation on the cell-level embedding feature through the second network layer.

[0038] In some embodiments, the hybrid expert network includes a shared expert network, a routing network, and a routing expert network. The step of extracting channel features of the radio environment based on the cell-level mask attention features through the hybrid expert network includes:

[0039] The shared expert network extracts shared features of the radio environment based on the cell-level mask attention features;

[0040] The routing expert networks are filtered through the routing network to obtain selected expert networks;

[0041] The selected expert network extracts expert features of the radio environment based on the cell-level mask attention features;

[0042] The shared features and the expert features are fused to obtain the channel features.

[0043] To achieve the above objectives, a second aspect of this application provides a radio environment sensing device, the device comprising:

[0044] An acquisition module is used to acquire a wireless measurement report of a radio environment; wherein the radio environment includes at least one cell, the cell has at least two beams, the beams include an unmasked beam and a masked beam, and the wireless measurement report includes a first measurement report of the unmasked beam and a second measurement report of the masked beam;

[0045] A tokenization processing module is used to tokenize the wireless measurement report to obtain a first token for the first measurement report and a second token for the second measurement report.

[0046] An input module is used to input the first token and the second token into a radio environment awareness model; wherein, the radio environment awareness model includes a backbone network, a prediction head and a projection head, the prediction head and the projection head are respectively connected to the backbone network, and the backbone network includes a hierarchical attention network and a hybrid expert network;

[0047] The attention calculation module is used to perform hierarchical attention calculation on the first token and the second token through the hierarchical attention network to obtain cell-level mask attention features;

[0048] An extraction module is used to extract channel features of the radio environment based on the cell-level mask attention features through the hybrid expert network;

[0049] The prediction module is used to predict the reference signal received power of the mask beam based on the channel characteristics using the prediction head, and to extract the projection embedding features of the mask beam based on the channel characteristics using the projection head.

[0050] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0051] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0052] The radio environment sensing method, radio environment sensing device, electronic device, and computer-readable storage medium proposed in this application overcome the limitations of CSI data acquisition by acquiring radio measurement reports of the radio environment and using these reports as the data foundation for a radio basic model. The radio measurement reports are tokenized, unifying heterogeneous reports into an unordered set of tokens to obtain a first token for the first measurement report and a second token for the second measurement report, thus solving the problems of variable data dimensions and semantic misalignment. The first and second tokens are input into the radio environment sensing model, which implicitly infers the radio propagation environment from a limited context. To address the sparsity and hierarchy of the radio measurement reports, a backbone network including a hierarchical attention network and a hybrid expert network is designed in the model. The hierarchical attention network performs hierarchical attention calculations on the first and second tokens to capture the signal propagation characteristics of the beam, obtaining cell-level masked attention features. The hybrid expert network further extracts features from the cell-level masked attention features to focus on the channel characteristics of the effective coherent region, obtaining the channel characteristics of the radio environment. By using a prediction head to predict the reference signal received power of the masked beam based on channel characteristics, and using a projection head to extract the projection embedding features of the masked beam based on channel characteristics, accurate perception of the radio environment is achieved, and the difficulty of radio environment perception is reduced. Attached Figure Description

[0053] Figure 1This is a flowchart of the radio environment sensing method provided in the embodiments of this application;

[0054] Figure 2 This is a schematic diagram of a wireless measurement report provided in an embodiment of this application;

[0055] Figure 3 This is an architecture diagram of the radio environment awareness model provided in the embodiments of this application;

[0056] Figure 4 This is a schematic diagram of the layered mask attention mechanism provided in an embodiment of this application;

[0057] Figure 5 yes Figure 1 The flowchart of step S140 in the middle;

[0058] Figure 6 This is a schematic diagram of a channel structure-guided hybrid expert provided in an embodiment of this application;

[0059] Figure 7 yes Figure 1 The flowchart of step S150 in the middle;

[0060] Figure 8 This is the strongest beam spatial distribution diagram provided in the embodiments of this application;

[0061] Figure 9 This is a flowchart of the training process of the radio environment perception model provided in the embodiments of this application;

[0062] Figure 10 This is a schematic diagram of the context self-distillation learning paradigm provided in the embodiments of this application;

[0063] Figure 11 yes Figure 9 The flowchart of step S920 in the middle;

[0064] Figure 12 yes Figure 9 The flowchart of step S930 in the text;

[0065] Figure 13 yes Figure 12 The flowchart of step S1210 in the middle;

[0066] Figure 14 This is a schematic diagram of the spatial partitioning of the training and test sets provided in an embodiment of this application;

[0067] Figure 15 This is a schematic diagram illustrating the impact of cell mask rate on RSRP completion error provided in an embodiment of this application;

[0068] Figure 16 This is a schematic diagram of the structure of the radio environment sensing device provided in the embodiments of this application;

[0069] Figure 17 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0071] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0072] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0073] With the development of sixth-generation (6G) wireless communication systems, various emerging application scenarios are constantly iterating, driving network performance indicators to become increasingly stringent and diversified. Against this backdrop, wireless digital twin technology based on artificial intelligence and deep learning can directly learn complex mapping relationships from observational data to simulate real-world wireless propagation environments that are difficult for traditional analysis methods to handle. Artificial intelligence methods are typically tailored to specific tasks and system configurations, requiring massive amounts of labeled data and exhibiting extremely poor generalization capabilities when deployed across different scenarios. When faced with new system configurations or channel conditions, it is often necessary to recollect data and retrain models, finely coordinating network parameters to adapt to the complex real-world environment. The convergence of heterogeneous technologies, network densification, and diversified service demands have dramatically increased the scale of network parameters and the complexity of their coupling with performance indicators. Therefore, the development of 6G communication is accelerating the transformation of network management towards data-driven intelligent self-optimization.

[0074] The emergence of Foundation Models (FMs) has revolutionized the field of artificial intelligence, acquiring extensive and generalizable knowledge through self-supervised pre-training on massive amounts of unlabeled data. The wireless communication field has also begun exploring wireless foundation models. Existing wireless foundation model technologies primarily focus on the physical layer, utilizing channel state information for tasks such as channel estimation, CSI feedback, and multi-user precoding. However, the physical layer heavily relies on high-fidelity CSI data, and collecting high-quality CSI data requires proprietary hardware and complex processing procedures. With the trend towards densification and large-scale multiple-input multiple-output (MIMO) wireless networks, the channel dimension is growing exponentially, making real-time acquisition of CSI data for all wireless links virtually impossible in practical engineering. This significantly increases the difficulty of radio environment awareness.

[0075] Based on this, embodiments of this application provide a radio environment sensing method, a radio environment sensing device, an electronic device, and a computer-readable storage medium, aiming to reduce the difficulty of radio environment sensing.

[0076] The radio environment sensing method, radio environment sensing device, electronic device, and computer-readable storage medium provided in the embodiments of this application are specifically described through the following embodiments. First, the radio environment sensing method in the embodiments of this application is described.

[0077] The radio environment sensing method provided in this application relates to the field of communications. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the radio environment sensing method, but is not limited to the above forms.

[0078] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0079] Consider a massive MIMO wireless communication system, where a cell is equipped with A uniform planar array of antennas This refers to the number of antennas in the horizontal direction. This represents the number of antennas in the vertical direction. The downtilt angle in free space. and azimuth They can be uniformly discretized into and From one angle. (The residential area) downlink channel impulse response from antenna to measurement location The model is as follows:

[0080] ,

[0081] in, for Channel power of the antenna angle path at the location; Antenna gain; The carrier wavelength; and These are the antenna spacings in the horizontal and vertical directions, respectively. and They represent Antenna at location and Random phase error of the antenna at the location.

[0082] The cell transmits M directional reference signal beams via beam scanning. Let the precoding matrix of the m-th beam be... Then, the reference signal received power (RSRP) of the m-th beam measured by the user equipment is... for:

[0083] ,

[0084] in, This refers to the base station's transmission power.

[0085] Based on the assumption of random phase error, the following statistical relationship exists between the multi-beam reference signal received power measurement vector rsrp and the channel angular power spectrum:

[0086] ,

[0087] Where E represents expectation; E[rsrp] represents expectation of rsrp; Represents the channel angular power spectrum; matrix The beamforming direction matrix represents the cell, and R represents the real number space; The total number of angles. .

[0088] Figure 1 This is an optional flowchart of the radio environment sensing method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S110 to S160.

[0089] Step S110: Obtain a wireless measurement report of the radio environment; wherein the radio environment includes at least one cell, the cell has at least two beams, the beams include an unmasked beam and a masked beam, and the wireless measurement report includes a first measurement report of the unmasked beam and a second measurement report of the masked beam.

[0090] Step S120: Tokenize the wireless measurement report to obtain a first token for the first measurement report and a second token for the second measurement report;

[0091] Step S130: Input the first token and the second token into the radio environment perception model; wherein, the radio environment perception model includes a backbone network, a prediction head and a projection head, the prediction head and the projection head are respectively connected to the backbone network, and the backbone network includes a hierarchical attention network and a hybrid expert network.

[0092] Step S140: Perform hierarchical attention calculation on the first token and the second token through a hierarchical attention network to obtain cell-level mask attention features;

[0093] Step S150: Extract channel features of the radio environment based on cell-level mask attention features using a hybrid expert network;

[0094] Step S160: The reference signal received power of the mask beam is predicted based on the channel characteristics by the prediction head, and the projection embedding features of the mask beam are extracted based on the channel characteristics by the projection head.

[0095] In step S110 of some embodiments, to overcome the problem that existing wireless infrastructure models rely on CSI data and have poor generalization ability, this application embodiment uses a wireless measurement report (MR) as the data basis. A wireless measurement report is a ubiquitous, continuously monitored, and automatically collected telecommunications big data set, containing reference signal received power for multiple cells and multiple beams, and containing real-world channel environment information. The wireless measurement report of the radio environment is obtained. The radio environment includes multiple cells, each cell has multiple beams, including unmasked beams and masked beams. The wireless measurement report includes a first measurement report for the unmasked beams and a second measurement report for the masked beams. The wireless measurement report includes the reference signal received power and physical configuration information of the beams. The reference signal received power is a continuous value, and the physical configuration information is the beam configuration information at the physical layer, including the beam radiation matrix, relative cell location, base station transmit power, carrier frequency, etc.

[0096] like Figure 2 As shown, Figure 2 Two MR data sets are provided: MR Sample 1 for Scenario 1 (network scenario 1) and MR Sample 2 for Scenario 2 (network scenario 2). The MR data were obtained from user equipment measurements. Both Scenario 1 and Scenario 2 deploy multiple base stations (BSs), which are divided into 4G and 5G base stations. The base stations cover multiple areas, including sector 1, sector 2, and sector 3. One or more sectors constitute a cell, and each cell is identified by a cell ID. Both MR Sample 1 and MR Sample 2 contain multiple cells, each with multiple beams. The beams include masked beams and unmasked beams; masked beams are identified by an "×".

[0097] from Figure 2 As can be seen from this, MR data has the following inherent characteristics: (1) missing location information; (2) a large number of missing values ​​in multi-beam RSRP, which has serious data sparsity; (3) the number of cells detected by different samples and the physical semantics are inconsistent, and the feature space is heterogeneous.

[0098] In step S120 of some embodiments, the heterogeneous wireless measurement reports are tokenized, and the wireless measurement reports are uniformly modeled as an unordered set of tokens to obtain the first token of the first measurement report and the second token of the second measurement report, which solves the problems of variable data dimensions and semantic misalignment.

[0099] For the m-th beam of cell q, as Figure 3 As shown, the token integrates both the RSRP Token and the instruction Token. The reference signal received power of the m-th beam in cell q is obtained through a multilayer perceptron. Projection as signal embedding Obtain the RSRP Token. Then, use the beam radiation matrix. The system aggregates physical configuration information such as the relative location of the cell and its transmission power to generate an embedded indicator. , and receive the instruction token.

[0100] For unmasked beams, the first token is represented as:

[0101] ,

[0102] For masked beams or missing beams, the RSRP Token is replaced with a learnable mask Token, which can be obtained through random initialization. The second token is represented as:

[0103] ,

[0104] in, This represents the Token mask.

[0105] Both the first token and the second token are input tokens.

[0106] The physical configuration information of cell q, such as the beam radiation matrix of the m-th beam, relative cell location, transmit power, and carrier frequency, is projected into a first embedding vector using a multilayer perceptron. A lookup table is then used to determine the second embedding vector corresponding to the cell type. The first and second embedding vectors are then fused to obtain the indication token. The cell type is a category obtained by classifying cells according to criteria such as coverage area, deployment scenario, and functional characteristics.

[0107] In step S130 of some embodiments, the first token of the unmasked beam and the second token of the masked beam in each cell are input into the radio environment awareness model. For example... Figure 3 As shown, the input tokens of M beams from N cells, namely cell 1, cell 2, ..., cell N-1 and cell N, are input into the radio environment perception model.

[0108] Due to the lack of location labels, severe heterogeneity (inconsistent dimensionality and physical semantics among different samples), and extreme sparsity (missing a large number of beam measurements) in MR data, existing model architectures and training paradigms cannot directly and effectively utilize this massive amount of low-cost data. This application's embodiments design a structure-aware model architecture based on the inherent characteristics of MR data. The radio environment awareness model of this application's embodiments includes a backbone network, a prediction head, and a projection head. The backbone network adopts a structure-aware Transformer model architecture, with the backbone being a Transformer backbone, and is connected to both the prediction head and the projection head. The prediction head is used to output the reconstructed RSRP of the masked beam, and the projection head is used to output the projected embedding of the masked beam.

[0109] The radio environment awareness model can be applied to intelligent management and optimization systems for wireless communication networks, specifically in the following areas:

[0110] (1) Data analysis server of core network or operation and maintenance center. By deploying pre-trained basic model, offline pre-training and continuous updating of model is carried out using MR big data collected from the entire network.

[0111] (2) Base station or RAN intelligent controller. By deploying a task-specific model that has been fine-tuned with a small number of samples, it is used for downstream tasks such as real-time network performance prediction, coverage optimization, and beam management.

[0112] To address the sparsity and hierarchical nature of MR data, a backbone network for a hybrid expert system was designed, incorporating a hierarchical masked attention mechanism and channel structure guidance. A hierarchical attention network was constructed based on the hierarchical masked attention mechanism, and a hybrid expert network was constructed based on the channel structure guidance mechanism.

[0113] To efficiently handle the hierarchical structure of beams nested within a cell, the hierarchical mask attention mechanism employs a local-to-global learning strategy. Please refer to [link / reference]. Figure 4 The hierarchical attention network consists of a first network layer, an average pooling layer, and a second network layer. The first network layer is a shallow Transformer, and the second network layer is a deep Transformer. The hierarchical mask attention mechanism has a query matrix and a key matrix, and is divided into intra-cell beam-level mask attention, cell-level beam aggregation, and inter-cell cell-level mask attention.

[0114] Please see Figure 5 In some embodiments, step S140 may include, but is not limited to, steps S510 to S530:

[0115] Step S510: Perform first attention calculation on the first token and the second token through the first network layer to obtain the beam-level mask attention features within the cell;

[0116] Step S520: Cell-level beam aggregation is performed on the beam-level mask attention features within the cell using an average pooling layer to obtain cell-level embedding features;

[0117] Step S530: Perform second attention calculation on the cell-level embedding features through the second network layer to obtain cell-level mask attention features.

[0118] In step S510 of some embodiments, the first network layer is a shallow network that introduces a beam-level masking matrix to limit the attention range to the same cell beam. The first token of the unmasked beam and the second token of the masked beam in each cell are input to the first network layer for first attention calculation, resulting in the intra-cell beam-level masked attention features. The formula for the first attention calculation is expressed as:

[0119] ,

[0120] in, K is the query matrix; K is the key matrix; V is the value matrix; T represents the transpose operation; Indicates the dimension of the query matrix; This represents the beam-level mask matrix.

[0121] It should be noted that the query matrix is ​​obtained by multiplying the first token, the second token, and the query weight of the first network layer; the key matrix is ​​obtained by multiplying the first token, the second token, and the key weight of the first network layer; and the value matrix is ​​obtained by multiplying the first token, the second token, and the value weight of the first network layer.

[0122] If the community The If each beam is an unmasked beam, then =0; if the cell The If each beam is a masked beam, then for This is to prevent information leakage from the masked beam.

[0123] In step S520 of some embodiments, the beam-level mask attention features within the cell are average-pooled using an average pooling layer. The pooling operation compresses the fine-grained beam features into cell-level embeddings, achieving cell-level beam aggregation and obtaining cell-level embedding features. The formula for average pooling is expressed as:

[0124] ,

[0125] in, Represents the cell-level embedding features of cell q; Indicates the number of beams in cell q; This represents the intra-cell beam-level mask attention feature of the m-th beam in cell q.

[0126] In step S530 of some embodiments, the second network layer is a deep network that introduces a cell-level mask matrix. It performs a second attention calculation on the aggregated cell-level embedding features to learn global cell dependencies and obtain cell-level mask attention features. The formula for the second attention calculation is expressed as:

[0127] ,

[0128] in, K is the query matrix; K is the key matrix; V is the value matrix; T represents the transpose operation; Indicates the dimension of the query matrix; This represents a cell-level mask matrix.

[0129] It should be noted that the query matrix is ​​obtained by multiplying the cell-level embedding features by the query weights of the second network layer, the key matrix is ​​obtained by multiplying the cell-level embedding features by the key weights of the second network layer, and the value matrix is ​​obtained by multiplying the cell-level embedding features by the value weights of the second network layer.

[0130] If the community and community If feature interaction is allowed between them, then =1; if the cell and community If feature interaction is not allowed between them, then It is 0.

[0131] Steps S510 to S530 above, based on the hierarchical nature of MR data, design a hierarchical attention mechanism to capture the local to global dependency relationship according to the intra-cell and inter-cell course learning method, thereby achieving network-wide wireless signal perception.

[0132] Please see Figure 6 The feedforward network of the Transformer model is replaced with a channel structure-guided hybrid expert network. This channel structure-guided hybrid expert network includes a shared expert network, a routing network, and multiple routing expert networks. Cell-level masked attention features can be used as input to extract channel features of the radio environment through the hybrid expert network.

[0133] Please see Figure 7 In some embodiments, step S150 may include, but is not limited to, steps S710 to S740:

[0134] Step S710: Extract shared features of the radio environment based on cell-level mask attention features through a shared expert network;

[0135] Step S720: Filter the routing expert networks through the routing network to obtain the selected expert networks;

[0136] Step S730: Extract expert features of the radio environment based on cell-level mask attention features by selecting an expert network;

[0137] Step S740: Perform feature fusion on the shared features and expert features to obtain the channel features.

[0138] In step S710 of some embodiments, the cell-level mask attention features are input into a shared expert network for feature extraction to obtain shared features of the radio environment.

[0139] In step S720 of some embodiments, the cell-level mask attention features are input to the routing network, and the routing network outputs the selection probability of the routing expert network. The routing expert network with the highest selection probability is selected as the expert network.

[0140] In step S730 of some embodiments, the expert features of the radio environment are obtained by selecting an expert network to extract features from the cell-level mask attention features. The expert features can be beam radiation maps.

[0141] In step S740 of some embodiments, the shared features and expert features are added together to obtain the channel features of the radio environment. The formula for extracting the channel features is expressed as:

[0142] ,

[0143] Where h represents the cell-level masked attention feature; MoE represents the channel feature; and SE represents the shared expert network. This indicates the selection of an expert network, where k is the expert index of the expert network with the highest selection probability output by the routing network. .

[0144] To ensure the expert network focuses on specific coherent channel regions, a strongest beam prediction auxiliary task is introduced, explicitly aligning the expert network's specialized representation with the angle of arrival of the physical propagation path. The expert index is the strongest beam index, and the beam radiation maps corresponding to the strongest beam indices at different locations, such as beam 5 radiation map and beam 1 radiation map, form the strongest beam spatial distribution, as shown in the image. Figure 8 As shown, the beam pattern covers both horizontal and downward tilt dimensions, effectively encoding the angle of arrival of the dominant propagation path.

[0145] The cross-entropy loss for the strongest beam prediction auxiliary task is expressed as:

[0146] ,

[0147] in, Cross-entropy loss; M is the number of routing expert networks; m is the expert index; For indicator functions, if If true, the indicator function outputs 1; otherwise, it outputs 0. Index of the true strongest beam; [M] represents the probability that m is the strongest beam index; [M] is the set of strongest beam indices.

[0148] Through the above steps S710 to S740, the channel characteristics of the radio environment can be extracted to perform radio environment perception based on the channel characteristics.

[0149] In step S160 of some embodiments, the reference signal received power of the mask beam is reconstructed based on the channel characteristics by the prediction head, and the channel characteristics are projected into the projection embedding characteristics of the mask beam by the projection head. The projection embedding characteristics are wireless signal representations, such as spatial representations (such as angle of arrival, horizontal angle, beam spatial distribution), frequency representations (such as channel frequency response, channel state information, etc.), and time representations (such as channel impulse response, multipath delay, etc.).

[0150] This application embodiment is based on the ubiquitous MR data base for model training. By utilizing unlabeled, highly sparse and heterogeneous feature space measurement report data, a network-level wireless basic model with physical structure perception capability is constructed to achieve cross-scenario general radio environment understanding and network-wide intelligent optimization.

[0151] Please see Figure 9 In some embodiments, the training process of the radio environment awareness model may include, but is not limited to, steps S910 to S940:

[0152] Step S910: Obtain the sample wireless measurement report; wherein, the sample wireless measurement report includes a first sample measurement report of the unmasked sample beam and a second sample measurement report of the masked sample beam, and the second sample measurement report includes the sample reference signal received power and physical configuration information of the masked sample beam;

[0153] Step S920: Wireless feature extraction is performed on the first sample measurement report and physical configuration information using a preset model to obtain global mask features, predicted reference signal received power of mask sample beams, and predicted projection embedding features.

[0154] Step S930: Calculate the loss based on the received power of the sample reference signal, global mask features, predicted received power of the reference signal, and predicted projection embedding features to obtain the target loss value.

[0155] Step S940: Update the preset model according to the target loss value to obtain the radio environment perception model.

[0156] In step S910 of some embodiments, this application embodiment transforms the utilization of MR data into a context-aware masked beam reconstruction task. Sample wireless measurement reports are acquired and divided into a disjoint context set and a target set. The context set consists of a first sample measurement report of the unmasked sample beam, and the target set consists of a second sample measurement report of the masked sample beam. The first sample measurement report includes the reference signal received power and physical configuration information of the unmasked sample beam. The second sample measurement report includes the sample reference signal received power and physical configuration information of the masked sample beam.

[0157] Let the dataset be Each MR sample is represented as an unordered set. ,in Let N represent the union of cells. For the community The number of beams, For the community Beam Reference signal received power, For the community Beam The physical configuration information. The first sample measurement report is represented as... The second sample measurement report is expressed as .

[0158] In step S920 of some embodiments, the first sample measurement report and physical configuration information are input into a preset model for wireless feature extraction to obtain global mask features, predicted reference signal received power of mask sample beams, and predicted projection embedding features.

[0159] In step S930 of some embodiments, a target loss value is calculated based on the sample reference signal received power, global mask features, predicted reference signal received power, and predicted projection embedding features. The target loss value is used to measure the expected prediction difference on the mask sample beam.

[0160] In step S940 of some embodiments, the target loss value is minimized, and the model parameters of the preset model are optimized to obtain a radio environment perception model. The optimization objective can be set as:

[0161] ,

[0162] Where E represents the expectation; Indicates a preset model; The model parameters are those of the preset model; the first sample measurement report is represented as follows. The second sample measurement report is expressed as follows: c and y represent the physical configuration information of the masked sample beam and the received power of the sample reference signal, respectively.

[0163] This optimization objective prompts the model to operate within a limited context. Implicit inference of the signal propagation environment allows for the understanding of underlying physical interaction patterns.

[0164] Through the above steps S910 to S940, we can get rid of the dependence on expensive CSI data and use massive, unlabeled MR data to train a radio environment perception model to establish a general understanding of the radio environment, thereby achieving zero-shot or few-shot network optimization across scenarios and multiple tasks.

[0165] Please see Figure 10 This application proposes a robust self-supervised learning framework based on contextual self-distillation. This framework employs a hybrid learning strategy, combining token-level mask reconstruction and feature-level self-distillation. Model training is performed based on multi-cell multi-beam RSRP using the contextual self-distillation learning paradigm. The contextual self-distillation learning paradigm is a two-branch pre-training framework combining generative reconstruction and discriminative alignment, introducing a student network (student encoder) and a teacher network (teacher encoder). The student and teacher networks have the same network architecture as the Transformer backbone described above.

[0166] The preset model includes a student encoder, a teacher encoder, a preset prediction head, a first projection head, and a second projection head. The student encoder is connected to both the preset prediction head and the first projection head, while the teacher encoder is connected to the second projection head. The network architecture of the first and second projection heads is identical. The student encoder is used to encode features of the mask version and configuration information of the multi-cell multi-beam RSRP, while the teacher encoder is used to encode features of the enhanced version and configuration information of the multi-cell multi-beam RSRP. The teacher encoder is updated by applying an exponential moving average (EMA) to the model parameters of the student encoder. The second projection head is updated by applying an exponential moving average to the model parameters of the first projection head.

[0167] Please see Figure 11 In some embodiments, step S920 may include, but is not limited to, steps S1110 to S1170:

[0168] Step S1110: The first sample measurement report and physical configuration information are feature-encoded by the student encoder to obtain the encoded features; wherein, the encoded features include global mask features and a first feature, the first feature being obtained based on the global mask features;

[0169] Step S1120: The received power of the prediction reference signal is output based on the first feature by a preset prediction head;

[0170] Step S1130: Output the first projection embedding feature based on the first feature using the first projection head;

[0171] Step S1140: Enhance the first sample measurement report and physical configuration information to obtain an enhanced view;

[0172] Step S1150: Encode the augmented view using the teacher encoder to obtain the second feature;

[0173] Step S1160: Output the second projection embedding feature based on the second feature using the second projection head;

[0174] Step S1170: Determine the predicted projection embedding feature based on the first projection embedding feature and the second projection embedding feature.

[0175] In step S1110 of some embodiments, the first sample measurement report of the unmasked sample beam and the physical configuration information of the masked sample beam are input to the student encoder for feature encoding to obtain encoded features. The encoded features include global mask features and a first feature. The network architecture of the student encoder is the same as the Transformer backbone. The global mask features are the global dependencies output by the hierarchical masking attention mechanism, and the first feature is the final feature output by the student encoder, obtained based on the global mask features.

[0176] In step S1120 of some embodiments, the reference signal received power of the mask sample beam is reconstructed based on the first feature by a preset prediction head to obtain the predicted reference signal received power.

[0177] In step S1130 of some embodiments, the first feature is mapped to a first projected embedding feature by the first projection head.

[0178] In step S1140 of some embodiments, to provide the teacher encoder with a globally compliant view, a channel model-driven data augmentation technique is introduced to enhance the first sample measurement report and physical configuration information, resulting in an enhanced view. The data augmentation technique implements two types of enhancement: power domain enhancement and beam domain enhancement. Power domain enhancement is used to simulate transmit power fluctuations and measurement noise, and is expressed as follows:

[0179] ,

[0180] ,

[0181] in, The received power of the reference signal included in the first sample measurement report; To implement the reference signal received power after power domain enhancement; and All are enhancement parameters; P is the transmit power; To implement the enhanced transmit power in the power domain.

[0182] Beam domain based on channel model Based on the statistical relationships, the sparse recovery algorithm is used to estimate the channel beamforming direction matrix. and reconstruct the full-beam RSRP This physically completes the missing beam, where q represents the q-th cell.

[0183] By simulating diverse physical propagation conditions through power domain enhancement and beam domain enhancement, the model can learn robust and generalizable wireless representations from massive amounts of unlabeled MR data.

[0184] In step S1150 of some embodiments, the difference between the student encoder and the teacher encoder is that the student encoder processes local context. The teacher encoder processes the global augmented view. The augmented view is feature-encoded using a teacher encoder to obtain the second feature.

[0185] In step S1160 of some embodiments, the second feature is mapped to a second projected embedding feature by the second projection head.

[0186] In step S1170 of some embodiments, the first projection embedding feature and the second projection embedding feature are used as predicted projection embedding features.

[0187] Through the above steps S1110 to S1170, the predicted projection embedding features can be obtained, and the loss can be calculated based on the predicted projection embedding features.

[0188] Please see Figure 12 In some embodiments, step S930 may include, but is not limited to, steps S1210 to S1250:

[0189] Step S1210: Calculate the first loss value based on the received power of the sample reference signal, the received power of the predicted reference signal, and the predicted projection embedding features;

[0190] Step S1220: Calculate the second loss value based on the predicted reference signal received power;

[0191] Step S1230: Perform the strongest beam prediction based on the global mask features to obtain the expert routing probability;

[0192] Step S1240: Calculate the third loss value based on the expert routing probability;

[0193] Step S1250: The first loss value, the second loss value, and the third loss value are weighted and calculated to obtain the target loss value.

[0194] In step S1210 of some embodiments, the predicted projection embedding feature includes a first projection embedding feature and a second projection embedding feature. A first loss value is obtained by calculating the loss based on the received power of the sample reference signal, the received power of the predicted reference signal, the first projection embedding feature, and the second projection embedding feature. The first loss value is a bi-target loss consisting of reconstruction loss and distillation loss.

[0195] In step S1220 of some embodiments, the MR data contains a large number of natively missing beams due to protocol mechanisms. These natively missing beams lack real labels. This application embodiment introduces a regularization loss function based on the physical prior that missing beams typically have weak signals. Let the known minimum RSRP of a cell within a certain MR sample be... Based on regularized loss function for missing beam sets RSRP prediction value A penalty is applied, resulting in a second loss value. The regularization loss function is expressed as:

[0196] ,

[0197] in, Indicates the second loss value; This represents the received power of the predicted reference signal for the masked sample beam (missing beam). For the missing beam set, it can be set as the mask sample beam set.

[0198] In step S1230 of some embodiments, referring to step S720, the global mask features output by the backbone network in the student encoder are input into the routing network in the student encoder to perform the strongest beam prediction and obtain the expert routing probability.

[0199] In step S1240 of some embodiments, the expert routing probability is substituted into the cross-entropy loss calculation formula to calculate the cross-entropy loss and obtain the third loss value.

[0200] In step S1250 of some embodiments, the overall optimization objective of the pre-training phase is a weighted sum of the bi-objective loss, the auxiliary routing loss, and the regularization loss. The target loss value is obtained by weighted summing of the first loss value, the second loss value, and the third loss value. The formula for calculating the target loss value is expressed as follows:

[0201] ,

[0202] in, Indicates the target loss value; This represents the first loss value; Indicates the second loss value; Indicates the third loss value; and All are weighting coefficients.

[0203] After pre-training, the model serves as a general-purpose wireless feature extractor. In practical applications such as network coverage prediction and zero-shot site deployment evaluation, excellent cross-city and cross-scenario generalization performance can be achieved simply by replacing the projection head with a lightweight task-specific network head and fine-tuning it with a very small amount of labeled data.

[0204] Through the above steps S1210 to S1250, the target loss value can be obtained, and the model parameters can be optimized based on the target loss value.

[0205] Please see Figure 13 In some embodiments, step S1210 may include, but is not limited to, steps S1310 to S1330:

[0206] Step S1310: Calculate the reconstruction loss based on the received power of the sample reference signal and the received power of the predicted reference signal;

[0207] Step S1320: Calculate the distillation loss based on the first projection embedding feature and the second projection embedding feature;

[0208] Step S1330: Calculate the first loss value based on the reconstruction loss and distillation loss.

[0209] In step S1310 of some embodiments, the difference between the received power of the sample reference signal and the received power of the predicted reference signal is calculated to obtain the reconstruction loss. The reconstruction loss is used to enable the student network to recover the RSRP value of the mask beam. The reconstruction loss is expressed as:

[0210] ,

[0211] in, Indicates the losses incurred during reconstruction; Indicates the student encoder. This represents the network parameters of the student encoder; Indicates the prediction head; This indicates the received power of the sample reference signal.

[0212] In step S1320 of some embodiments, the first projected embedding features are mapped to a probability distribution to obtain the student distribution. The second projected embedding features are mapped to a probability distribution to obtain the teacher distribution. Cross-entropy is calculated based on the student and teacher distributions to obtain the distillation loss. The distillation loss is expressed as:

[0213] ,

[0214] in, This represents the distillation loss; K represents the number of feature dimensions; k represents the k-th feature dimension; Indicates the distribution of teachers; This indicates the distribution of students.

[0215] Student distribution is defined as:

[0216] ,

[0217] in, Temperature coefficient; The first projection embedding feature; K is the number of feature dimensions; and k respectively represent the first The kth feature dimension.

[0218] The teacher distribution introduces a centralization bias. and a lower temperature coefficient This is to prevent the model from collapsing.

[0219] In step S1330 of some embodiments, the reconstruction loss and distillation loss are weighted and summed to obtain a first loss value. The first loss value is expressed as:

[0220] ,

[0221] in, This is the first loss value; To enhance the view.

[0222] Through the above steps S1310 to S1330, the dual-target loss can be obtained, and the model parameters can be optimized based on the dual-target loss to improve the wireless signal extraction performance of the model.

[0223] This application proposes a network-level wireless fundamental model driven by massive measurement report data, aiming to overcome the heavy reliance of existing physical layer models on expensive and scarce channel state information. Considering the lack of location information, extreme sparsity, and heterogeneous feature space of real MR data, this application makes systematic innovations in data utilization, model architecture, and learning paradigm. The innovations are as follows:

[0224] (1) The embodiments of this application transform the utilization of unlabeled MR data into a context-aware mask beam reconstruction task, and construct a unified input sequence by fusing beam-level RSRP tags and physical configuration instruction tags, so as to overcome the heterogeneous feature space and variable dimension problem of MR data by using this tagging mechanism.

[0225] (2) Structure-aware Transformer backbone network architecture design. Two core components conforming to physical structure priors are introduced: a hierarchical masking attention mechanism and a channel structure-guided hybrid expert system. The hierarchical masking attention mechanism adopts a progressive feature extraction method from intra-cell beam-level masking attention to inter-cell cell-level aggregation and attention; the channel structure-guided hybrid expert system uses the strongest beam prediction as an auxiliary task to guide the routing network, enabling different expert networks to focus on specific coherent channel regions.

[0226] (3) Context-based self-distillation self-supervised pre-training paradigm. This training paradigm combines a dual-objective pre-training framework of generative reconstruction and discriminative alignment. The framework employs a joint optimization strategy that combines local mask reconstruction loss and momentum teacher network global feature distillation loss, a channel model-driven data augmentation method (power domain augmentation and beam domain augmentation based on sparse recovery), and a physical regularization loss strategy introduced for the natively missing beams in MR data.

[0227] The advantages of this application's embodiments are mainly reflected in the following aspects: (1) Extremely low data acquisition and generalization costs. It abandons the reliance on high-cost drive testing and dedicated hardware CSI collection, and directly utilizes ubiquitous, automatically collected unlabeled MR data in the network. Through self-supervised pre-training, the model can learn general radio representations. When facing new system configurations or base station deployments, it only needs a very small amount of label data for fine-tuning (Few-shot) or even zero-shot to achieve cross-city and cross-scenario generalization, which greatly reduces the manpower and time costs of network optimization; (2) It has powerful heterogeneous data processing and physical consistency modeling capabilities. The introduction of input tokenization and hierarchical mask attention mechanism solves the problems of variable MR data dimensions and severe missing data; (3) Combining channel structure-guided hybrid expert systems and physical-driven data augmentation, the model can implicitly infer complex radio propagation environments from the context. Compared with traditional task-specific models, it has stronger physical interpretability, generalization ability and robustness.

[0228] The performance of the method in this application embodiment is evaluated based on a large-scale real-world MR dataset. The MR dataset used comes from a real network, covering a 10km × 8km urban area, and includes 952 cells operating in the 2.1GHz and 3.5GHz frequency bands. This dataset contains 1.6 million MR samples.

[0229] To evaluate the generalization ability of the model, such as Figure 14 As shown, the dataset is strictly divided into geographical regions to obtain training and test sets, where (X,Y) represents the location coordinates of the geographical region, with X as the x-axis and Y as the y-axis. The training set contains 1,281,075 training samples from 738 cells, and the test set contains 289,685 test samples from the remaining 214 cells. This division ensures that the training and test sets cover completely different and non-overlapping geographical regions, effectively validating the model's adaptability to unknown environments. The mean absolute error (MAE) of the model on the test set for the multi-beam RSRP completion task of masked cells is used as the performance metric to measure its zero-shot generalization ability in novel scenarios.

[0230] With a cell mask rate of 40%, ablation experiments were conducted on the method of this embodiment, and the results are shown in Table 1. During the ablation experiments, additional components were added to each step based on the previous step. The experiments show that each component designed in this embodiment contributes a positive gain to the model's generalization ability. Based on the synergistic effect of these components, the zero-sample MAE of the model was reduced to 5.32 dB, demonstrating the advancement and rationality of the overall model framework design.

[0231] Table 1

[0232]

[0233] Furthermore, this application's embodiments also analyze the impact of cell mask ratio (cell mask rate) on model performance and distinguish the completion errors of the primary serving cell and neighboring cells. For example... Figure 15 As shown, with the increase in the cell mask ratio, the MAE of both the main server cell and neighboring cells steadily increases, which aligns with the intuitive expectation of increasing task difficulty. Meanwhile, from... Figure 15It was observed that the completion MAE of neighboring cells was consistently higher than that of the primary serving cell. This phenomenon is consistent with physical laws. The primary serving cell has a stronger radio signal, its propagation path is relatively predictable, and it is more robust to noise. In contrast, the radio signals from neighboring cells are usually weaker and more susceptible to interference from complex terrain such as canyons, multipath reflections, and diffraction effects, making their behavior patterns difficult to model accurately. Even so, the model in this application embodiment can still control the completion error of neighboring cells to within 6dB even with a cell masking rate as high as 60%, fully demonstrating its strong robustness under extremely high data missing rates.

[0234] The downstream fine-tuning capability of the pre-trained pedestal model was tested on MR localization. The model achieved a localization error of 44.87m on 80% of the training samples, which is better than the 61.48m localization performance obtained by retraining the model.

[0235] Please see Figure 16 This application also provides a radio environment sensing device that can implement the above-described radio environment sensing method. The radio environment sensing device includes:

[0236] The acquisition module 1610 is used to acquire a wireless measurement report of a radio environment; wherein the radio environment includes at least one cell, the cell has at least two beams, the beams include an unmasked beam and a masked beam, and the wireless measurement report includes a first measurement report of the unmasked beam and a second measurement report of the masked beam;

[0237] The tokenization processing module 1620 is used to tokenize the wireless measurement report to obtain a first token for the first measurement report and a second token for the second measurement report.

[0238] The input module 1630 is used to input the first token and the second token into the radio environment perception model; wherein, the radio environment perception model includes a backbone network, a prediction head and a projection head, the prediction head and the projection head are respectively connected to the backbone network, and the backbone network includes a hierarchical attention network and a hybrid expert network;

[0239] The attention calculation module 1640 is used to perform hierarchical attention calculation on the first token and the second token through a hierarchical attention network to obtain cell-level mask attention features.

[0240] Extraction module 1650 is used to extract channel features of the radio environment based on cell-level mask attention features through a hybrid expert network;

[0241] The prediction module 1660 is used to predict the reference signal received power of the mask beam based on channel characteristics using the prediction head, and to extract the projection embedding features of the mask beam based on channel characteristics using the projection head.

[0242] The specific implementation of this radio environment sensing device is basically the same as the specific embodiment of the radio environment sensing method described above, and will not be repeated here.

[0243] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described radio environment sensing method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0244] Please see Figure 17 , Figure 17 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0245] The processor 1710 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0246] The memory 1720 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1720 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1720 and is called and executed by the processor 1710 to execute the radio environment sensing method of the embodiments of this application.

[0247] The input / output interface 1730 is used to implement information input and output;

[0248] The communication interface 1740 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0249] Bus 1750 transmits information between various components of the device (e.g., processor 1710, memory 1720, input / output interface 1730, and communication interface 1740);

[0250] The processor 1710, memory 1720, input / output interface 1730 and communication interface 1740 are connected to each other within the device via bus 1750.

[0251] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described radio environment sensing method.

[0252] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0253] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0254] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0255] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0256] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0257] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0258] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0259] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0260] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0261] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0262] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0263] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A radio environment sensing method, characterized in that, The method includes: Obtain a wireless measurement report of a radio environment; wherein the radio environment includes at least one cell having at least two beams, the beams including an unmasked beam and a masked beam, and the wireless measurement report includes a first measurement report of the unmasked beam and a second measurement report of the masked beam; The wireless measurement report is tokenized to obtain a first token for the first measurement report and a second token for the second measurement report; The first token and the second token are input into the radio environment awareness model; wherein, the radio environment awareness model includes a backbone network, a prediction head and a projection head, the prediction head and the projection head are respectively connected to the backbone network, and the backbone network includes a hierarchical attention network and a hybrid expert network; The first token and the second token are subjected to hierarchical attention calculation through the hierarchical attention network to obtain cell-level mask attention features; The hybrid expert network extracts channel features of the radio environment based on the cell-level mask attention features; The prediction head predicts the reference signal received power of the masked beam based on the channel characteristics, and the projection head extracts the projection embedding features of the masked beam based on the channel characteristics.

2. The method according to claim 1, characterized in that, The radio environment awareness model is trained according to the following steps: Obtain a sample wireless measurement report; wherein, the sample wireless measurement report includes a first sample measurement report of an unmasked sample beam and a second sample measurement report of a masked sample beam, the second sample measurement report including the sample reference signal received power and physical configuration information of the masked sample beam; Wireless feature extraction is performed on the first sample measurement report and the physical configuration information using a preset model to obtain global mask features, the predicted reference signal received power of the mask sample beam, and the predicted projection embedding features. The target loss value is obtained by calculating the loss based on the received power of the sample reference signal, the global mask feature, the received power of the predicted reference signal, and the predicted projection embedding feature; The preset model is updated based on the target loss value to obtain the radio environment perception model.

3. The method according to claim 2, characterized in that, The preset model includes a student encoder, a teacher encoder, a preset prediction head, a first projection head, and a second projection head. The student encoder is connected to both the preset prediction head and the first projection head, and the teacher encoder is connected to the second projection head. The step of performing wireless feature extraction on the first sample measurement report and the physical configuration information using the preset model to obtain global mask features, the predicted reference signal received power of the mask sample beam, and the predicted projection embedding features includes: The student encoder performs feature encoding on the first sample measurement report and the physical configuration information to obtain encoded features; wherein, the encoded features include the global mask feature and a first feature, the first feature being obtained based on the global mask feature; The preset prediction head outputs the received power of the prediction reference signal based on the first feature; The first projection head outputs a first projection embedding feature based on the first feature; The first sample measurement report and the physical configuration information are enhanced to obtain an enhanced view; The enhanced view is feature-encoded using the teacher encoder to obtain the second feature; The second projection head outputs a second projection embedding feature based on the second feature; The predicted projection embedding feature is determined based on the first projection embedding feature and the second projection embedding feature.

4. The method according to claim 2, characterized in that, The step of calculating the target loss value based on the received power of the sample reference signal, the global mask features, the received power of the predicted reference signal, and the predicted projection embedding features includes: A first loss value is calculated based on the received power of the sample reference signal, the received power of the predicted reference signal, and the predicted projection embedding features; Calculate the second loss value based on the predicted reference signal received power; Based on the global mask features, the strongest beam prediction is performed to obtain the expert routing probability; Calculate the third loss value based on the expert routing probability; The target loss value is obtained by weighting the first loss value, the second loss value, and the third loss value.

5. The method according to claim 4, characterized in that, The predicted projection embedding feature includes a first projection embedding feature and a second projection embedding feature. The method is based on the received power of the sample reference signal, the received power of the predicted reference signal, and the predicted projection embedding feature. Calculate the first loss value, including: The reconstruction loss is calculated based on the received power of the sample reference signal and the received power of the predicted reference signal; The distillation loss is calculated based on the first projection embedding feature and the second projection embedding feature; The first loss value is calculated based on the reconstruction loss and the distillation loss.

6. The method according to any one of claims 1 to 5, characterized in that, The hierarchical attention network includes a first network layer, an average pooling layer, and a second network layer. The hierarchical attention calculation performed on the first token and the second token through the hierarchical attention network to obtain cell-level mask attention features includes: The first attention calculation is performed on the first token and the second token by the first network layer to obtain the beam-level mask attention features in the cell. The average pooling layer is used to perform cell-level beam aggregation on the beam-level mask attention features within the cell to obtain cell-level embedding features. The cell-level mask attention feature is obtained by performing a second attention calculation on the cell-level embedding feature through the second network layer.

7. The method according to any one of claims 1 to 5, characterized in that, The hybrid expert network includes a shared expert network, a routing network, and a routing expert network. The step of extracting channel features of the radio environment based on the cell-level mask attention features using the hybrid expert network includes: The shared expert network extracts shared features of the radio environment based on the cell-level mask attention features; The routing expert networks are filtered through the routing network to obtain selected expert networks; The selected expert network extracts expert features of the radio environment based on the cell-level mask attention features; The shared features and the expert features are fused to obtain the channel features.

8. A radio environment sensing device, characterized in that, The device includes: An acquisition module is used to acquire a wireless measurement report of a radio environment; wherein the radio environment includes at least one cell, the cell has at least two beams, the beams include an unmasked beam and a masked beam, and the wireless measurement report includes a first measurement report of the unmasked beam and a second measurement report of the masked beam; A tokenization processing module is used to tokenize the wireless measurement report to obtain a first token for the first measurement report and a second token for the second measurement report. An input module is used to input the first token and the second token into a radio environment awareness model; wherein, the radio environment awareness model includes a backbone network, a prediction head and a projection head, the prediction head and the projection head are respectively connected to the backbone network, and the backbone network includes a hierarchical attention network and a hybrid expert network; The attention calculation module is used to perform hierarchical attention calculation on the first token and the second token through the hierarchical attention network to obtain cell-level mask attention features; An extraction module is used to extract channel features of the radio environment based on the cell-level mask attention features through the hybrid expert network; The prediction module is used to predict the reference signal received power of the mask beam based on the channel characteristics using the prediction head, and to extract the projection embedding features of the mask beam based on the channel characteristics using the projection head.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images