Communication-aided perception wireless infrastructure model and construction method and application thereof

Abstract

The present application relates to a communication auxiliary perception wireless base model and its construction method and application, the construction method comprises: S1: the original channel state information data is converted into radio frequency input embedding sequence;S2: the context dependence relationship in the radio frequency input embedding sequence is learned by using a multilayer bidirectional encoder, and the deep coding is completed;S3: a part of the component unit in the radio frequency input embedding sequence is randomly masked, and the pre-training is carried out in the manner of predicting the local embedding vector mapped by the masked component unit, and the model construction is completed.The model is constructed by using the construction method.The application is the application of the model in the communication auxiliary perception task.The model of the present application is pre-trained by self-supervision on a large amount of unlabeled channel state information data, learns in multiple dimensions, excavates the space-time-frequency characteristics of the wireless channel, obtains the channel knowledge, and is fine-tuned by the labeled sample for the downstream task, which can realize high-precision and high-robustness communication auxiliary perception.

Description

Technical Field

[0001] This invention relates to a wireless basic model for communication-assisted sensing, its construction method, and its application, belonging to the field of channel model technology. Background Technology

[0002] With the continuous evolution of wireless communication technology, the demand for intelligent development of 6G networks is constantly increasing. Traditional wireless signal processing methods based on task-specific designs are no longer sufficient to meet the service requirements of "intelligent and inclusive" services in complex environments. In recent years, drawing on the development ideas of foundation models in the field of artificial intelligence, the concept of Wireless Foundation Model (WFM) has been gradually proposed in the field of wireless communication. It aims to achieve unified representation modeling and capability transfer across tasks and scenarios through large-scale wireless data pre-training. Existing wireless systems usually rely on model structures designed for specific tasks (such as modulation recognition, channel estimation, or interference detection), which have limited generalization capabilities and require retraining or complex parameter adjustments when transferring to different scenarios, resulting in high system development costs and insufficient adaptability. In contrast, the Wireless Foundation Model can learn universal wireless signal representations through pre-training on large-scale radio frequency data, thereby achieving rapid adaptation in downstream tasks with a small number of samples.

[0003] However, applying existing large-scale wireless models to practical communication-assisted sensing tasks presents several interconnected and mutually restrictive limitations. First, the scarcity of high-quality wireless sensing data severely impacts the expressive power of the basic wireless model. Unlike regression tasks in mainstream communication tasks (such as channel prediction, beam prediction, and networking), training large-scale wireless sensing models relies on a large amount of labeled wireless sensing data. Due to the complexity of electromagnetic wave propagation, accurately labeling non-intuitive radio frequency signals is extremely difficult, typically requiring sophisticated specialized equipment or a large number of professionals for on-site surveys and data collection. This results in extremely high costs and limited scale for acquiring high-quality labeled data. Second, existing spatiotemporal-frequency feature learning methods are relatively simplistic and difficult to transfer to complex communication-assisted sensing classification tasks. Large-scale wireless sensing models not only need to learn generalized channel representations to ensure system generalization capabilities but also need to further explore the adaptability of channel features to flexibly adapt to diverse sensing tasks. However, current methods lack hierarchical encoding of multi-dimensional information when learning wireless channel representations, leading to a significant decrease in the performance of acquired channel features across scenarios and tasks in communication-assisted sensing. Finally, training large wireless sensing models (WSLMs) under the few-shot learning paradigm still faces significant technical challenges in reducing system deployment costs. Due to multipath propagation effects, radio frequency (RF) signals are highly sensitive to dynamic changes in the environment, making it difficult to effectively train high-accuracy task-specific classification models using only a small amount of labeled wireless sensing data. Meanwhile, interference noise, incomplete or inaccurate labels, channel fading, and distortion, which are prevalent in complex real-world environments, further exacerbate uncertainties and performance degradation during model training. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of existing technologies, this invention provides a wireless foundation model for communication-assisted sensing, its construction method, and its application. The wireless foundation model is self-supervised pre-trained on massive amounts of unlabeled channel state information data, learns using multi-dimensional representations, mines the spatiotemporal-frequency characteristics of wireless channels, acquires channel knowledge, and is fine-tuned using labeled samples for different downstream tasks, thereby achieving high-precision and highly robust communication-assisted sensing.

[0005] The technical solution adopted in this invention is: a method for constructing a wireless basic model for communication-assisted sensing, comprising the following steps: S1: Convert the raw channel state information data into an RF input embedding sequence; S2: A multi-layer bidirectional encoder is used as the core network architecture to learn the contextual dependencies in the radio frequency input embedding sequence, complete the deep encoding of radio frequency features, and obtain a general representation (feature representation) containing rich contextual information. S3: The model is pre-trained by using a dynamic masking method that randomly masks some components in the radio frequency input embedding sequence and predicts the local embedding vector of the masked component based on the context of the component. This method learns a general and robust (more robust) representation and completes the model construction.

[0006] Preferably, in step S1, the original channel state information data is organized into a sequence of radio frequency (RF) words. Each RF word is the amplitude or phase vector of all subcarriers at a single antenna and a single time interval. Multiple RF words within a continuous time interval constitute a radio frequency sequence. One-dimensional convolution and max-pooling operations are performed on each RF word to map it into a local embedding vector, i.e., a RF term. The input RF sequence is encapsulated into a sum vector containing term embeddings, segment embeddings, and position embeddings to construct a radio frequency input embedding sequence.

[0007] Preferably, in S2, the encoder uses a multi-head self-attention mechanism to capture the correlation between transmission distances inherent in the radio frequency input embedding sequence, modeling the spatial and temporal dependencies in the wireless channel.

[0008] Preferably, the encoder utilizes a multi-head self-attention mechanism to capture information in parallel from different representation subspaces.

[0009] Preferably, the multi-head self-attention mechanism calculates attention weights by aggregating contextual information through scaling dot products.

[0010] Preferably, in S3, the constituent units that need to be occluded are randomly selected again in each training round.

[0011] Furthermore, the constituent units are radio frequency words.

[0012] Preferably, in step S3, in the initial stage of training, a mask rate higher than a preset standard value is used for training, and as training progresses, the mask rate is linearly reduced to the preset standard value.

[0013] The wireless basic model for communication-assisted sensing is constructed using any of the wireless basic model construction methods for communication-assisted sensing disclosed in this invention.

[0014] The application of the wireless basic model for communication-assisted sensing disclosed in this invention in communication-assisted sensing tasks.

[0015] Preferably, when performing a communication-assisted sensing task, the wireless underlying model of the communication-assisted sensing is fine-tuned, including: Task-specific feature adaptation: The output of the pre-trained encoder on multiple antennas is globally max-pooled to obtain the feature vectors of each antenna and multi-head attention. The feature vectors of all antennas are concatenated to form a spatial feature map. Based on the dimension of the feature map, the convolution kernel size is adaptively selected by establishing a non-linear relationship between the dimension of the feature map and the receptive field, and features adapted to the current task are automatically extracted. Quality-guided perception task classification: Cluster analysis is performed on unlabeled pre-trained radio frequency data. By analyzing the eigenvalue gaps of the graph Laplacian matrix, the optimal number of clusters for the current dataset is automatically estimated. K-means clustering is performed on the samples, and the silhouette coefficient of each sample is calculated. The normalized silhouette coefficient is used as a weighting factor and multiplied with the adaptively extracted features specific to the task. The perception task classifier is trained using small sample labeled data to fine-tune the model, so that the wireless basic model can be adapted to specific communication-assisted perception tasks to achieve high-precision perception.

[0016] The beneficial effects of this invention are: (1) This invention learns general channel knowledge through BERT pre-training and then fine-tunes it for specific tasks to achieve high-precision and high-robust perception tasks under small sample conditions. It solves the problems of existing deep fingerprint localization methods relying on large-scale labeled data and poor environmental adaptability. It fundamentally improves the universality and environmental adaptability of signal feature representation and provides a brand-new solution for communication-assisted perception. (2) This invention uses multi-dimensional representation learning to deeply understand the spatiotemporal correlation of wireless channels from three levels: local (frequency domain), nearest neighbor (time domain) and global (spatial domain) of channel state information. It can effectively represent the channel characteristics under complex multipath propagation. At the same time, it can automatically extract task adaptive features through cross-antenna channel interaction, which significantly improves training efficiency. (3) By introducing clusterability quantization, the present invention enables the model parameters to adaptively align with specific tasks, and trains the task classifier with a small number of samples for fine-tuning, thereby effectively reducing the deployment cost of fully practical communication-assisted perception tasks. Attached Figure Description

[0017] Figure 1 This is a flowchart of the wireless basic model construction method for communication-assisted sensing of the present invention; Figure 2 This is an architecture diagram of the wireless basic model for communication-assisted sensing of the present invention; Figure 3 This is a flowchart of a communication-assisted sensing task performed using the wireless basic model of the communication-assisted sensing of this invention; Figure 4 This is a cumulative distribution function graph of complex indoor scene positioning using the present invention; Figure 5This is a cumulative distribution function graph for complex dynamic scene localization using the present invention. Detailed Implementation

[0018] This invention relates to a wireless foundational model for communication-assisted sensing, its construction method, and its applications. By employing a pre-trained model based on the BERT architecture and utilizing a self-attention mechanism, a dynamic mask training strategy is designed to learn the radio frequency context dependencies inherent in massive channel state information (CSI) data, thereby revealing the complex propagation characteristics of wireless channels and acquiring channel knowledge. Then, the Radio-BERT encoder is directly transferred to downstream sensing tasks. Global channel representations are obtained by aggregating deep features from multiple antennas, and task-specific adaptive features are extracted through cross-antenna interaction. Finally, a data quality-guided task classifier fine-tuning method is employed. This method utilizes small-sample labeled data combined with clusterability quantization weights to train the task classifier, addressing the performance degradation problem caused by insufficient data to support effective model learning.

[0019] See Figure 1 This invention discloses a method for constructing a wireless basic model for communication-assisted sensing, specifically including the following: 1. RF input embedding generation First, the raw channel state information data is organized into a sequence of radio frequency words. Specifically, the amplitude or phase vector of all subcarriers at a single antenna and a single moment is located as a radio frequency word. Multiple video words in a continuous time period constitute a radio frequency sentence, i.e. a radio frequency sequence. By learning the distance-related contextual information contained in the radio sequence, the channel changes caused by multipath propagation, path loss and shadow fading can be characterized.

[0020] Building upon this, to construct an input embedding suitable for Transformer model processing, the input radio frequency sequence is further encapsulated into a sum vector containing the following three embeddings: (1) Lexical embedding: For each radio frequency (RF) word, perform one-dimensional convolution and max pooling operations to map it into a local embedding vector, i.e., the RF word element. Let the original RF word vector be... ,in For the number of subcarriers, If the set is real numbers, then the computation of lexical embeddings can be expressed as: , in, A one-dimensional convolution kernel, For bias, These are the generated radio frequency (RF) terms. Additionally, special terms can be added at the beginning of the sequence to capture global scene features, and special terms can be added between sequences to separate different RF sequences.

[0021] (2) Segment embedding: To achieve synchronized processing of multiple radio sequences, segmented embeddings are learned. Each segmentation is identified by its corresponding radio sequence. During Radio-BERT pre-training, segmented embedding vectors provide the model with explicit sequence identifiers, enabling it to learn the relationships between pairs of radio sequences.

[0022] (3) Position embedding: A location embedding vector is introduced to provide distance-related information for each radio-frequency word in the input embedding sequence. The location embedding vector is initially a random vector with the same dimension as the word segment embedding vector and the segment embedding vector. During pre-training, when the Radio-BERT model calculates the association representation of each radio-frequency word through the self-attention mechanism, it can combine the location embedding vector to consider the distance-related association between words.

[0023] For the first in the sequence Each radio frequency term has a final input embedding representation that is the sum of the three embeddings mentioned above: , in, For the first The input embedding sequence of radio frequency terms, For the first The word embedding vector of each radio frequency word, For the first Segment embedding vectors of radio frequency words, For the first The position embedding vector of each radio frequency term, This represents the total number of embeddings.

[0024] In this way, the original, structurally complex channel state information data can be transformed into a model-understandable input radio frequency embedding sequence containing rich information.

[0025] 2. Transformer-based bidirectional RF encoder A multi-layer bidirectional Transformer encoder is employed as the core network architecture to learn deep contextual dependencies in the RF input embedding sequence. This encoder utilizes a multi-head self-attention mechanism to capture the inherent correlations between transmission distances within the RF input embedding sequence, thereby understanding the dependencies of the RF spatiotemporal context, effectively establishing a fundamental representation model of the wireless channel, and revealing the physical propagation characteristics of the wireless channel.

[0026] For the input embedded sequence , For radio frequency terms, Given the dimension of the radio frequency sequence, a self-attention mechanism is used to compute the query matrix. Key matrix Value matrix : , , , in, , , These are the learnable query projection matrix, key-value projection matrix, and value projection matrix, respectively. They embed the input... Mapping to the query space, it learns "how to query"; mapping to the key-value space, it learns "how to be queried"; and mapping to the value space, it learns "what information to extract". Furthermore... , , and These represent the query dimension, key dimension, and value dimension, respectively.

[0027] The self-attention mechanism aggregates contextual information by scaling the dot product to calculate attention weights, and its output is: , Employing a multi-head self-attention mechanism enables the model to capture information in parallel from different representation subspaces. , in, , For the number of attention heads, To output the projection matrix, The dimension of multi-head attention projection.

[0028] To further enhance the modeling capability for complex wireless propagation characteristics, the deep encoding of radio frequency features is finally completed through feedforward networks, residual connections, and layer normalization, resulting in a general feature representation containing rich contextual information.

[0029] 3. Dynamic mask pre-training To enable the model to learn general and robust representations (with higher robustness), this invention employs a dynamic masking pre-training strategy. During pre-training, a portion of the radio frequency words in the radio frequency input embedding sequence are randomly masked. The model predicts these masked words based on their context. To further improve the model's generalization ability and pre-training efficiency, a dynamic masking mechanism is introduced: (1) The mask position changes dynamically: In each training epoch, the positions of the words to be masked are randomly selected again instead of using a fixed masking pattern. This forces the model to learn a deeper understanding of the context rather than simply memorizing a fixed masking pattern, effectively preventing overfitting.

[0030] (2) Dynamic mask learning: Dynamic sequences generated using random masks This prevents the model from overfitting due to a fixed mask sequence, where the set of dynamic mask positions is... The output of the Radio-BERT encoder ,in These are the parameters that the encoder can learn. The implicit embeddings learned by the Radio-BERT encoder. The estimated probability is obtained by applying softmax: , in, These are masked word elements in a dynamic sequence. w These are model weights. b It is model bias. For each element in the set of dynamic mask locations, the training loss function for the Radio-BERT encoder is: , (3) Dynamic adjustment of mask rate: A linearly decreasing mask rate scheduling strategy is employed. In the early stages of training, a higher mask rate (above a preset standard value) is used. This introduces greater reconstruction difficulty and gradient smoothness, prompting the model to learn more robust global features. As training progresses, the mask rate linearly decreases to a lower standard value (the preset standard value) to facilitate fine-grained model convergence and improve training efficiency. This dynamic mask rate balances the efficiency and effectiveness of pre-training. The linear reduction scheduling strategy is expressed as: , in, For the first mask rate of the step, The initial mask rate, Total training steps This is the final mask rate.

[0031] Through the above dynamic mask pre-training, the model can gain a deep understanding of the spatiotemporal context of the wireless channel and learn a general wireless signal representation that is insensitive to noise, interference and environmental changes.

[0032] This invention also discloses a wireless foundation model for communication-assisted sensing, which is constructed using the wireless foundation model construction method for communication-assisted sensing disclosed in this invention. Its model architecture is as follows: Figure 2 As shown.

[0033] See Figure 3This invention also discloses the application of the wireless foundation model for communication-assisted sensing in communication-assisted sensing tasks. For downstream communication-assisted sensing tasks, it fully utilizes the powerful channel representation and generalization capabilities of Radio-BERT to fine-tune the model, achieving accurate communication-assisted sensing. The spatiotemporal correlation representation of multiple antennas is directly obtained using the trained Radio-BERT encoder, and the channel features of multiple antennas are aggregated to acquire global channel knowledge. Then, cross-antenna interaction is used to extract task-adaptive features. Finally, a quality-guided task classification method is employed, quantifying the clusterability of small sample data to jointly train a task-specific classifier.

[0034] (1) Task-specific feature adaptation: Extracting the RF spatiotemporal correlation representation of a single antenna using a pre-trained Radio-BERT encoder For the channel representation matrix Global max pooling is performed, and the feature vector is obtained by multi-head attention aggregation. In multi-antenna communication systems, the channel feature vector is expanded into a spatial feature map. Its dimensions , It refers to the number of antennas in the array. This refers to the number of heads that attract attention. In practice... The dimensionality of a spatial channel determines its spatiotemporal complexity. High-dimensional feature maps require long-range interactive representations of global channel characteristics, and vice versa. Furthermore, low-dimensional feature maps are prone to introducing noise due to excessively large receptive fields. This invention adaptively selects the convolution kernel size and extracts task-adaptive features by mapping the receptive field to the spatial feature map dimension.

[0035] Based on feature map dimensions The size of the one-dimensional convolution kernel is adaptively selected through a nonlinear mapping. : , in, , This function establishes a non-linear relationship between feature map dimensionality and receptive field, enabling the network to dynamically adjust the coverage of the convolutional kernel based on the complexity and dimensionality of the input data. This adaptive mechanism avoids the problems of insufficient feature extraction or overfitting caused by using a fixed kernel size. (The last sentence appears to be incomplete and possibly refers to further details about reusing the kernel.) The values ​​are used to perform convolutional feature extraction on the CSI embedding to obtain task-adaptive features.

[0036] (2) Quality-guided prediction based on clusterability assessment: During the fine-tuning phase, due to the limited number of labeled samples, data quality (such as noise or interference) can severely impact training results. To address this issue, a quality-guided task-specific classification method is employed. This method quantifies the clusterability of data to guide model fine-tuning training, enabling the acquisition of an accurate communication-assisted perception task classifier through small-sample learning.

[0037] ①Estimation of the number of clusters in wireless data: The optimal number of clusters is estimated using the characteristic gaps of spectral clustering. K Specific algorithm steps: Step 1: Construct a similarity graph. For a given wireless dataset... Construct an undirected graph with each sample as a vertex. ,in It is the set of all vertices. It is the set of edges connecting any two vertices. The similarity between any two samples is calculated based on the Gaussian radial basis function. The similarity matrix is ​​obtained. .

[0038] , in, and This represents different wireless data samples.

[0039] Step 2: Construct the Graph Laplacian matrix Calculate the degree matrix. The Graph Laplace matrix is ​​obtained. .right Perform symmetric normalization ,in, It is an identity matrix.

[0040] Step 3: Calculation The eigenvalues ​​of . Solve The eigenvalues ​​are then arranged in ascending order: .

[0041] Step 4: Find the largest feature gap and determine the optimal feature gap. Calculate the difference between adjacent eigenvalues. Find the Largest index : , ② Quantification of clusterability of wireless data: Wireless dataset Perform K-means clustering, with the number of clusters being... Given a sample ,calculate The average distance to all other samples within its cluster, i.e. Cohesion for: , in, yes Distance from other samples, A smaller value indicates a denser cluster. Simultaneously, the calculation... Separation For the sample Average distance to all samples in every other cluster: , in, and Indicates the number of samples in two different clusters. Reflects the sample The distance from the nearest non-clustered element; the larger this value, the farther the sample is from other clusters. Calculate the data sample. contour coefficient The following formula is used to measure the matching degree between the sample and its assigned cluster (compared to other clusters): .

[0042] During fine-tuning training, the normalized silhouette coefficients are... As a weighting factor, it is multiplied by the adaptively extracted features and then input into the classifier for training. This operation allows the model to use only samples with high silhouette coefficients (close to 1) when updating parameters, whose features are more representative and discriminative. A small amount of labeled CSI data was selected from the pre-training dataset to construct the fine-tuning dataset: , It is CSI data. The tag, This refers to the number of samples in the fine-tuning dataset. (Task-specific classifier) Loss function for training: , in, It is cross-entropy. Through the joint fine-tuning of task-specific feature adaptation and quality-guided classification training, this invention can efficiently adapt a general wireless basic model to specific communication-assisted sensing tasks using a few-shot learning paradigm, achieving high-precision sensing performance.

[0043] Unless otherwise specified or further limited to one preferred or optional technical means being another, the preferred and optional technical means disclosed in this invention can be arbitrarily combined to form several different technical solutions.

[0044] Comparative example: See Figure 4 and Figure 5 When using existing WiFi signals and employing the communication-assisted sensing method of this invention for positioning in complex indoor scenes, the positioning accuracy is 1.36m; for positioning in complex dynamic scenes, the robust positioning RMSE error is 1.46m. In contrast, using traditional deep learning methods for communication-assisted sensing, the positioning accuracy in complex indoor scenes is 2.75m; for positioning in complex dynamic scenes, the robust positioning RMSE error is 2.47m. The comparison shows that, compared to traditional deep learning methods for communication-assisted sensing, the method of this invention improves positioning accuracy by 50.5% in complex indoor scenes and improves robust positioning RMSE error by 40.9% in complex dynamic scenes.

Claims

1. A method for constructing a wireless basic model for communication-assisted sensing, characterized in that... Includes the following steps: S1: Convert the raw channel state information data into an RF input embedding sequence; S2: A multi-layer bidirectional encoder is used as the core network architecture to learn the contextual dependencies in the radio frequency input embedding sequence, complete the deep encoding of radio frequency features, and obtain a general representation containing rich contextual information. S3: The model is pre-trained by using a dynamic masking method that randomly masks some components in the radio frequency input embedding sequence and predicts the local embedding vector of the masked component based on the context of the component. This method learns a general and robust representation and completes the model construction.

2. The method for constructing a wireless basic model for communication-assisted sensing according to claim 1, characterized in that... In S1, the original channel state information data is organized into a radio frequency word sequence. A radio frequency word is the amplitude or phase vector of all subcarriers at a single antenna and a single moment. Multiple radio frequency words in a continuous time period constitute a radio frequency sequence. The input radio frequency sequence is encapsulated into a sum vector containing word embedding, segment embedding and position embedding to construct a radio frequency input embedding sequence.

3. The method for constructing a wireless basic model for communication-assisted sensing according to claim 1, characterized in that... In S2, the encoder uses a multi-head self-attention mechanism to capture the correlation between transmission distances inherent in the RF input embedding sequence, modeling the spatial and temporal dependencies in the wireless channel.

4. The method for constructing a wireless basic model for communication-assisted sensing according to claim 3, characterized in that... The encoder uses a multi-head self-attention mechanism to capture information in parallel from different representation subspaces.

5. The method for constructing a wireless basic model for communication-assisted sensing according to claim 3, characterized in that... The multi-head self-attention mechanism calculates attention weights by scaling the dot product and aggregating contextual information.

6. The method for constructing a wireless basic model for communication-assisted sensing according to claim 1, characterized in that... In S3, during each training round, the constituent units that need to be occluded are randomly selected again.

7. The method for constructing a wireless basic model for communication-assisted sensing according to claim 1, characterized in that... In S3, during the initial training phase, a mask rate higher than the preset standard value is used for training. As training progresses, the mask rate is linearly reduced to the preset standard value.

8. A wireless basic model for communication-assisted sensing, characterized in that... The wireless basic model is constructed using the communication-assisted sensing wireless basic model construction method described in any one of claims 1-7.

9. The application of the wireless basic model of communication-assisted sensing as described in claim 8 in communication-assisted sensing tasks.

10. The application of the wireless basic model for communication-assisted sensing according to claim 9, characterized in that... Fine-tuning the wireless underlying model of the communication-assisted sensing includes: Task-specific feature adaptation: The output of the pre-trained encoder on multiple antennas is globally max-pooled to obtain the feature vectors of each antenna and multi-head attention. The feature vectors of all antennas are concatenated to form a spatial feature map. Based on the dimension of the feature map, the convolution kernel size is adaptively selected by establishing a non-linear relationship between the dimension of the feature map and the receptive field, and features adapted to the current task are automatically extracted. Quality-guided prediction: Cluster analysis is performed on unlabeled pre-trained radio frequency data. By analyzing the eigenvalue gaps of the graph Laplacian matrix, the optimal number of clusters for the current dataset is automatically estimated. K-means clustering is performed on the samples, and the silhouette coefficient of each sample is calculated. The normalized silhouette coefficient is used as a weighting factor and multiplied with the adaptively extracted features specific to the task. The perception task classifier is trained using small sample labeled data to fine-tune the model, so that the wireless basic model can be adapted to specific communication-assisted perception tasks to achieve high-precision perception.

Images