An unmanned aerial vehicle air-to-ground information sharing integrated communication method based on semantic communication

Abstract

This invention discloses a UAV-based air-to-ground integrated sensing communication method, belonging to the field of UAV communication and sensing integration technology. The method includes the following steps: constructing a three-dimensional model of a UAV-assisted ISAC system; the ISAC system three-dimensional model includes a ground station, a UAV, and the communication link between the ground station and the UAV; analyzing the Channel Identity Components (CSI) in the ISAC system three-dimensional model to obtain channel characteristic information in the time domain, spatial domain, and Doppler domain; formulating a CSI feedback scheme based on the obtained CSI characteristic information; and verifying the formulated CSI feedback method. This invention generates a channel subspace using the AE algorithm and the K-means algorithm to determine the optimal number of semantic database elements; and uses the K-means algorithm to cluster the low-dimensional features obtained by the encoder, improving clustering performance and ensuring high-precision performance of the ISAC task while reducing feedback load; further verification of the high-precision performance of the method for the ISAC task is achieved through blocking detection, beam prediction, and scene recognition.

Description

Technical Field

[0001] This invention belongs to the field of integrated communication and sensing technology for unmanned aerial vehicles (UAVs), specifically relating to an integrated air-to-ground communication method for UAVs based on semantic communication. Background Technology

[0002] In sixth-generation communication networks, unmanned aerial vehicle (UAV)-assisted communication (UAV-AEC) is expected to play a crucial role in integrated air-space-terrestrial-sea communication. Meanwhile, Information-Sensing Communication (ISAC) technology has garnered significant attention due to its reliable and efficient UAV-AEC capabilities. In ISAC systems, transceivers are no longer limited to data transmission; they also collect environmental information, providing data support for various applications. Unlike terrestrial communication, UAV communication possesses unique attributes, including three-dimensional (3D) scattering space, three-dimensional velocity, three-dimensional rotation, and three-dimensional antenna arrays, all of which significantly impact channel characteristics. Furthermore, UAV-AEC systems face power and energy limitations because UAV operation relies on limited batteries or fuel, necessitating improvements to make ISAC systems more suitable for UAV-AEC systems.

[0003] In ISAC tasks such as congestion detection, beam prediction, and scene recognition, the transmitter requires Channel State Information (CSI). Therefore, the CSI is estimated at the receiver and then sent back to the transmitter. In the context of scene recognition, some research has been conducted on converting channel characteristics into physical propagation scenarios. However, transmitting complete CSI information increases the energy consumption of the UAV, thereby reducing the efficiency of UAV-assisted ISAC systems.

[0004] Recent literature has proposed superimposed CSI feedback schemes based on neural networks, novel CSI sensing and recovery networks using deep learning, etc. However, these feedback schemes cannot meet the high real-time requirements of UAV communication and do not take into account relevant channel theory.

[0005] Therefore, how to improve the accuracy of ISAC tasks while meeting the real-time and low-energy requirements of UAV-assisted ISAC systems is the key technical problem that this invention aims to solve. Summary of the Invention

[0006] The purpose of this invention is to provide a semantic communication-based air-to-ground integrated communication method for unmanned aerial vehicles (UAVs) to solve the problems mentioned in the background art.

[0007] The objective of this invention is achieved as follows: a semantic communication-based air-to-ground integrated communication method for unmanned aerial vehicles (UAVs), characterized by the following steps:

[0008] Step S1: Construct a 3D model of the UAV-assisted ISAC system;

[0009] The 3D model of the ISAC system includes a ground station, a drone, and the communication link between the ground station and the drone.

[0010] Step S2: Analyze the CSI in the three-dimensional model of the ISAC system to obtain channel characteristic information in the time domain, spatial domain, and Doppler domain;

[0011] Step S3: Based on the acquired CSI characteristic information, formulate a CSI feedback plan;

[0012] Step S4: Verify the established CSI feedback method.

[0013] Preferably, the communication link between the UAV and the ground station is affected by distance, obstacles, and the speed of the UAV, and these effects are reflected by CSI; the CSI describes the channel behavior in the time delay domain through CIR, expressed as:

[0014]

[0015] Where, α l For each path, the complex amplitude, τ l Delay for each path, θ l Horizontal arrival angle, φ for each path l Vertical angle of arrival and f for each path d Let L be the Doppler frequency shift for each path, L be the total number of multipaths, δ(t) be the impulse function, and j be the imaginary unit.

[0016] Preferably, in step S3, a CSI feedback scheme is formulated based on the acquired CSI characteristic information, and the specific operation is as follows:

[0017] Step S3-1: Construct a CSI semantic database based on deep clustering;

[0018] Step S3-2: Identify the CSI semantic database;

[0019] Step S3-3: Construct and generate a feedback scheme based on the CSI semantic database and recognition results.

[0020] Preferably, the construction of the CSI semantic database based on deep clustering in step S3-1 includes the following steps:

[0021] Step S3-1-1: Use channel samples H = {h1,…,h} M CSI generates a semantic library, where h m ,1≤m≤M is the CIR of the m-th channel, where m is the number of samples;

[0022] Step S3-1-2: Generate the channel subspace using the AE algorithm and the K-means algorithm. The AE algorithm includes an encoder and a decoder; the encoder compresses the input channel data into low-dimensional features.

[0023] Z = f(W) e H+b e );

[0024] Where Z is the encoded feature, W e and b e Here are the encoder's weight matrix and bias vector; f(·) is the activation function; the AE loss in the network is expressed as:

[0025]

[0026] The decoder takes the compressed feature Z as input and reconstructs data that is as similar as possible to the original data:

[0027] H′=f(W d Z+b2);

[0028] Where H′ is the reconstructed data, W d and b d These are the weight matrix and bias vector of the decoder;

[0029] Step S3-1-3: Use K-means to cluster the low-dimensional features obtained by the encoder.

[0030] Preferably, in step S3-1-3, K-means is used to cluster the low-dimensional features obtained by the encoder. The specific operation is as follows:

[0031] First, randomly select k data points from the channel dataset as initial centroids, and calculate the distance between each data point and all centroids; then assign each data point to the nearest centroid to form k clusters.

[0032] For each cluster, a new centroid is calculated by averaging all data points assigned to that cluster, and the updated centroid is compared with the previous centroid.

[0033] If the centroids do not change significantly, or the maximum number of iterations has been reached, the algorithm stops; otherwise, the process is repeated. Once the algorithm converges, it returns the final centroids and cluster assignments. The core metric is the sum of squared errors (SSE).

[0034]

[0035] Where, m k It is a semantic database C i The center.

[0036] Preferably, the AE algorithm is used for feature learning of channel data, and the K-means function is used for clustering features. The clustering loss of the K-means function is expressed as:

[0037]

[0038] Among them, s ik This indicates whether a channel sample belongs to the k-th semantic database; 1 indicates belonging and 0 indicates not belonging.

[0039] By updating the loss function, joint training of feature learning and clustering processes is achieved:

[0040] min L=αL n +βL c , α≥0, β>0;

[0041] Here, α and β are weighting factors.

[0042] Preferably, in step S3-2, the CSI semantic database is identified, and the incremental K-means algorithm is used to identify the channel subspace. The specific operation is as follows:

[0043] Step S3-2-1: Input the new channel data h into the trained autoencoder to generate two features, and calculate the distance between the new channel data and all K current cluster centers;

[0044] Step S3-2-2: Assign the new channel data to the cluster represented by the nearest cluster center, and update the learning rate parameter of the selected cluster center. The learning rate parameter is:

[0045] C new =C old +α*(hC old );

[0046] Among them, C old It is the old cluster centroid, C new It is the new cluster centroid; α ranges from (0,1) and is used to control the magnitude of the cluster centroid update;

[0047] Step S3-2-3: Compare the distance between the new and old clusters with the threshold μ; if the result is less than the threshold, the new channel data belongs to the cluster; if the result is greater than the threshold, it means that a new channel subspace structure has emerged, and all channel data need to be re-clustered.

[0048] Preferably, the verification of the CSI feedback method in step S4 includes using the silhouette coefficient S and F measure to evaluate the clustering performance and verifying the wave velocity prediction and scene recognition in the CSI feedback scheme.

[0049] The silhouette coefficient S is used to compare the performance of different clustering algorithms and depends only on the actual distribution of the clusters; the silhouette coefficient S is defined as follows:

[0050]

[0051] Where a(x) represents the average distance between x and MPCs belonging to the same cluster, and b(x) represents the minimum distance between x and MPCs belonging to different clusters; S is in the range of [-1,1], and the larger the value, the better the clustering performance;

[0052] The F-measure is a robustness metric used to balance precision and recall; the F-measure is defined as follows:

[0053]

[0054] Where PB and RB represent the average of precision and recall, and F is in the range of [-1,1]. The larger the value, the better the clustering robustness.

[0055] Compared with the prior art, the present invention has the following improvements and advantages:

[0056] 1. The channel subspace is generated by the AE algorithm and the K-means algorithm to determine the optimal number of semantic databases; and the K-means algorithm is used to cluster the low-dimensional features obtained by the encoder to improve the clustering performance, ensuring the high accuracy of the ISAC task, while greatly reducing the feedback load.

[0057] 2. Through verification of blockage detection, beam prediction and scene recognition, the method proposed in this invention improves the accuracy of the three ISAC tasks and significantly reduces the amount of feedback information; the high F-value and silhouette coefficient S obtained by deep clustering demonstrate their effectiveness in capturing the underlying structure of CSI and achieving more accurate clustering results. Attached Figure Description

[0058] Figure 1 This is the overall flowchart of the present invention.

[0059] Figure 2 A framework diagram of the ISAC system for drone assistance.

[0060] Figure 3 A city scene map with specific feature parameters.

[0061] Figure 4 A CDF comparison chart of the silhouette coefficient S for deep clustering and traditional K-means.

[0062] Figure 5 A comparison chart of CDF for deep clustering and the traditional K-means F measure.

[0063] Figure 6 This is a model diagram of the beam prediction scheme.

[0064] Figure 7 This is a comparison chart of ROC results for different methods for blockage detection and beam prediction.

[0065] Figure 8 This is a diagram of a scene recognition solution.

[0066] Figure 9 This is a confusion matrix diagram for scene recognition based on traditional methods.

[0067] Figure 10 This is a confusion matrix diagram for scene recognition based on semantic feedback. Detailed Implementation

[0068] The invention will be further summarized below with reference to the accompanying drawings.

[0069] like Figure 1 As shown, a semantic communication-based UAV air-to-ground integrated sensing communication method includes the following steps:

[0070] Step S1: Construct a 3D model of the UAV-assisted ISAC system;

[0071] like Figure 2 As shown, the 3D model of the ISAC system includes a ground station, a drone, and the communication link between the ground station and the drone. It is assumed that the ground station acts as a receiver, estimates the original CSI of the air-to-ground (A2G) channel and performs semantic database recognition, and the drone server acts as a transmitter, performing different tasks based on the CSI feedback.

[0072] At the ground station, the estimated CSI (denoted as h) u The initial extraction was performed by a trained unsupervised artificial network autoencoder (AE), which extracted data from h. u Extract z u It plays a crucial role in CSI feedback schemes; by leveraging its encoding and decoding capabilities, the AE aims to capture the underlying structure and features of CSI without requiring labeled data. Using a trained AE from h u Extract z u Then, incremental clustering is used to connect the current channel with the semantic database c. u Matching is performed; incremental clustering is a dynamic identification method that updates the existing semantic database upon receiving new CSIs. This method enables the ISAC system to dynamically and accurately assign the current channel to the appropriate semantic database. In the case of UAVs, the ISAC system associates the current channel with the matched semantic database to extract the CSIs required for the ISAC task.

[0073] Step S2: Analyze the CSI in the three-dimensional model of the ISAC system to obtain channel characteristic information in the time domain, spatial domain, and Doppler domain;

[0074] The communication link between the human-machine interface and the ground station is affected by distance, obstacles, and the speed of the drone, and these effects are reflected by the Channel Indicator (CSI). The CSI describes the channel's behavior in the time-delay domain using the Channel Indicator (CIR), expressed as follows:

[0075]

[0076] Where, α l For each path, the complex amplitude, τ l Delay for each path, θ l Horizontal arrival angle, φ for each path l Vertical angle of arrival and f for each path d Let L be the Doppler shift for each path, L be the total number of multipaths, δ(t) be the impulse function, and j be the imaginary unit. Analysis of the CIR (Channel Irregular Response) yields channel characteristics in the time, spatial, and Doppler domains, which play a crucial role in various ISAC (Inter-Area Compression) tasks.

[0077] For large-scale fading, path loss is a commonly used metric to describe signal power attenuation. Path loss refers to the decrease in signal strength that occurs as the distance between the UAV and the ground station increases, and can be expressed as:

[0078] PL(dB)=P T (dBm)-P R (dBm);

[0079] Among them, P T and P R These are the transmit power and receive power. P R Calculated using the following formula:

[0080]

[0081] The Rice factor is a commonly used metric for describing small-scale fading. It describes the power ratio of the maximum power path to other paths. A larger Rice factor indicates the possible presence of a strong LOS component, meaning that the direct path from the UAV to the ground station dominates the received signal power. The Rice factor can be calculated as follows:

[0082]

[0083] in, The amplitude of the main multipath component.

[0084] In the time delay domain, the root mean square delay spread (RMS-DS) is typically used to characterize channel dispersion caused by small-scale fading and multipath effects, representing the degree of delay spread in the channel. The definition of RMS-DS can be expressed as:

[0085]

[0086] Where, τ l and α l It refers to the delay and amplitude of each propagation path. The average time delay can be expressed as:

[0087]

[0088] In the spatial domain, angular spread (AS) is commonly used to describe the magnitude of angular dispersion and is widely applied in beamforming and antenna array design for UAV communication systems. The AS of the azimuth of arrival can be expressed as:

[0089]

[0090] Where, θ l It is the azimuth of arrival for each propagation path. It is the average angle, which can be expressed as:

[0091]

[0092] Furthermore, if θ l Replace with elevation angle It is possible to obtain the AS at the elevation angle.

[0093] In the Doppler domain, the Doppler effect is used to describe the frequency shift of the received signal caused by the motion of the UAV.

[0094]

[0095] Where v is the velocity of the UAV, λ is the wavelength, and θ is the velocity of the UAV. d It is the angle between the direction of motion of the radio wave and the direction of incident.

[0096] Step S3: Based on the acquired CSI characteristic information, formulate a CSI feedback plan;

[0097] Step S3-1: Construct a CSI semantic database based on deep clustering;

[0098] Step S3-1-1: Construct a CSI semantic database using deep clustering technology, and learn the underlying structure and patterns of CSI samples using deep learning technology; CSI samples are represented as:

[0099] H = {h1,h2,…,h}M};

[0100] Where M is the number of preset samples; using a deep clustering method combining AE and K-means, the preset channel samples H can be divided into K disjoint subspaces, satisfying the following conditions:

[0101]

[0102] Among them, C i (1≤i≤K) is the i-th semantic database; clustering channel samples into the semantic database provides a systematic and effective method for storing and retrieving relevant CSIs; the semantic database organizes channel data in a way that simplifies the retrieval of information required for a specific ISAC task, and stores representative labels associated with each dataset, which capture basic information about the corresponding channel features.

[0103] Step S3-1-2: Generate the channel subspace using the AE algorithm and the K-means algorithm. The AE algorithm includes an encoder and a decoder; the encoder compresses the input channel data into low-dimensional features.

[0104] Z = f(W) e H+b e );

[0105] Where Z is the encoded feature, W e and b e is the encoder's weight matrix and bias vector; f(·) is the activation function;

[0106] The decoder takes the compressed feature Z as input and reconstructs data that is as similar as possible to the original data:

[0107] H′=f(W d Z+b2);

[0108] Where H′ is the reconstructed data, W d and b d These are the weight matrix and bias vector of the decoder;

[0109] The loss of an effect (AE) in a network is represented as:

[0110]

[0111] Step S3-1-3: Cluster the low-dimensional features obtained by the encoder using K-means;

[0112] First, randomly select k data points from the channel dataset as initial centroids, and calculate the distance between each data point and all centroids; then assign each data point to the nearest centroid to form k clusters.

[0113] For each cluster, a new centroid is calculated by averaging all data points assigned to that cluster, and the updated centroid is compared with the previous centroid.

[0114] If the centroids do not change significantly, or the maximum number of iterations has been reached, the algorithm stops; otherwise, the process is repeated. Once the algorithm converges, it returns the final centroids and cluster assignments. The core metric is the sum of squared errors (SSE).

[0115]

[0116] Where, m k It is a semantic database C i The center of the data. As the number of clusters K increases, the sample partitioning becomes more refined, and the degree of clustering within each cluster gradually increases, leading to a natural decrease in SSE. Furthermore, when K is less than the actual number of clusters, increasing K significantly increases the degree of clustering within each cluster, causing a sharp drop in SSE. However, when K reaches the actual number of clusters, any further increase in K will produce diminishing returns in terms of clustering, causing the rate of SSE decrease to drop sharply, and then as K continues to increase, the rate of SSE decrease gradually levels off. Therefore, the value of K represents the actual number of clusters in the data.

[0117] The AE algorithm is used for feature learning of channel data, and the K-means function is used for clustering features. Joint training of feature learning and clustering processes is achieved by simultaneously updating the AE algorithm and the K-means function.

[0118] The clustering loss of the K-means function is expressed as:

[0119]

[0120] Among them, s ik This indicates whether a channel sample belongs to the k-th semantic database; 1 indicates belonging and 0 indicates not belonging.

[0121] By updating the loss function, joint training of feature learning and clustering processes can be achieved, as shown below:

[0122] min L=αL n +βL c , α≥0, β>0;

[0123] Here, α and β are weighting factors.

[0124] Step S3-2: Identify the CSI semantic database;

[0125] Channel subspace identification is a crucial step in the CIR feedback scheme. UAVs have diverse application scenarios, and channel data from almost all of these scenarios cannot be fed back into the channel subspace. Therefore, the identification algorithm first needs to determine whether the current UAV communication channel belongs to the generated channel subspace. If it does, it needs to accurately identify the corresponding channel sub-region; otherwise, it needs to be re-clustered.

[0126] Incremental K-means algorithm is used to identify the channel subspace. Incremental K-means is an extension of traditional clustering algorithms, designed to capture the constantly changing patterns and structures in dynamic channel datasets. It combines the principles of subspace identification with the adaptive capabilities of evolutionary clustering to effectively track changes in the channel subspace. When the UAV receives new channel data, if the data does not deviate from historical expectations, it needs to be matched into the existing cluster. However, if the data structure has changed significantly, the clustering results must be modified to reflect the new structure. When new data points arrive, the following operations are performed:

[0127] Step S3-2-1: Input the new channel data h into the trained autoencoder to generate two features, and calculate the distance between the new channel data and all K current cluster centers;

[0128] Step S3-2-2: Assign the new channel data to the cluster represented by the nearest cluster center, and update the learning rate parameter of the selected cluster center. The learning rate parameter is:

[0129] C new =C old +α*(hC old );

[0130] Among them, C old It is the old cluster centroid, C new It is the new cluster centroid; α ranges from (0,1) and is used to control the magnitude of the cluster centroid update;

[0131] Step S3-2-3: Compare the distance between the new and old clusters with the threshold μ; if the result is less than the threshold, the new channel data belongs to the cluster; if the result is greater than the threshold, it means that a new channel subspace structure has emerged, and all channel data need to be re-clustered.

[0132] Step S3-3: Construct and generate a feedback scheme based on the CSI semantic database and recognition results.

[0133] Taking an A2G drone system in an urban environment as an example, this demonstrates offline training and online deployment; the dataset is the foundation for semantic database generation and recognition in an urban environment, and its details are as follows... Figure 3As shown; to capture the complete channel characteristics of the UAV A2G system in urban scenarios, we adopted a standard urban digital map according to R-REC-P.1410, which defines three parameters to represent the urban digital map: α0 represents the proportion of building area to total area, β0 represents the average number of buildings per unit area (number of buildings per square kilometer), and γ0 represents the building height that follows a Rayleigh distribution.

[0134]

[0135] To cover as many urban scenes as possible, the value of α0 was set to 0.1 to 0.8 in 0.1 intervals, β0 was set to 100 to 800 in 100 intervals, and γ0 was set to 8 to 50 in 2 intervals. In the 176 generated geophysical scenes, 1000 channel features were randomly obtained in each scene based on the ray tracing algorithm.

[0136] By generating a large number of channel samples in each city scenario, a wide range of channel conditions and variables are obtained; a semantic database is generated from the collected channel samples; based on the SSE trend, the optimal number of semantic databases can be determined when K=20.

[0137] Step S4: Verify the established CSI feedback method.

[0138] like Figure 4 , 5 As shown, the silhouette coefficient S and the F measure are used to evaluate clustering performance. The silhouette coefficient S can be used to compare the performance of different clustering algorithms, and it depends only on the actual distribution of the clusters. The definition of S is as follows:

[0139]

[0140] Where a(x) represents the average distance between x and MPCs belonging to the same cluster, and b(x) represents the minimum distance between x and MPCs belonging to different clusters; S is in the range of [-1,1], and the larger the value, the better the clustering performance;

[0141] The F-measure is a robustness metric used to balance precision and recall; the F-measure is defined as follows:

[0142]

[0143] Where PB and RB represent the average of precision and recall. F is in the range of [-1,1], and the larger the value, the better the clustering robustness.

[0144] The method proposed in this invention is compared and analyzed with several other methods. The specific process is as follows:

[0145] Step 1: Beam Prediction

[0146] Beam prediction is primarily used to improve communication quality by redirecting the beam to a more favorable target when the direct path is blocked; for example... Figure 6 As shown, in beam prediction based on the CSI semantic database and the feedback scheme generated from the recognition results, two important problems need to be solved: block prediction and target scatterer prediction.

[0147] In NLOS scenarios, predicting the beam direction to a more favorable direction by detecting target scatterers with maximum power is crucial. Besides channel characteristics, environmental features also play a significant role in beam prediction in NLOS scenarios. This paper compares the performance of beam prediction based on CSI semantic database feedback with methods in the literature. Traditional methods use raw CSI, while this invention uses statistical CSI obtained from a constructed semantic database. For blocking prediction, SVM is used as the classification method, which can accurately identify blocks or obstacles in the environment. Utilizing GNN for beam prediction is beneficial for capturing complex dependencies and relationships in graph-structured data. Furthermore, the datasets used for training and testing consist of samples with varying numbers of scatterers to ensure the generalization of the prediction method.

[0148] The receiver operating characteristic (ROC) curves for the blocking detection and beam prediction results obtained by the two methods are shown below. Figure 7 As shown in the figure, the ROC curve shows the true positive rate on the y-axis and the false positive rate on the x-axis. A higher true positive rate and a lower false positive rate indicate better classification performance. It can be seen that the system improves performance by 1% in blocking detection and by 2% in beam prediction. Simulation results demonstrate that the communication system based on the semantic feedback method exhibits high accuracy and effectiveness in both detection and beam prediction, ensuring reliable and efficient communication performance of the UAV-assisted ISAC system.

[0149] Step 2: Scene Recognition

[0150] like Figure 8 As shown, scene recognition is an important ISAC task in UAV communication to improve communication performance and optimize communication strategies; its goal is to accurately classify the current communication scene into these predefined classes in order to improve the system's performance and adaptability in different environments.

[0151] This paper compares scene recognition based on semantic feedback with traditional methods proposed in the literature. Traditional methods use channel features from the original CSI, while this invention uses statistical CSI channel features obtained from a semantic database. Both methods employ multi-class support vector machines, where the multi-class classifier is constructed from a binary support vector machine-based classifier. A total of 3000 channel samples were simulated for each scene. Of these, 2000 channel samples were used to train the SVM model, and the remaining 1000 channel samples were used for testing and performance evaluation.

[0152] Table 1 shows the specific simulation parameters and scene model parameters for suburban and three urban scenarios.

[0153]

[0154]

[0155] Confusion matrices are useful tools for evaluating the effectiveness of identification methods, providing detailed classifications of test data and actual labels. For example... Figure 9 , 10 As shown in the confusion matrix, both methods achieved 100% accuracy in both maritime and suburban scenarios. This verifies that the dynamic recognition method can gradually incorporate CSI related to maritime scenarios into the semantic database, even if this information may not initially exist in the predefined semantic database. The recognition accuracy differs by only 1% in urban, dense urban, and high-rise urban scenarios, indicating consistent performance across different urban environments. The overall recognition accuracy of the two methods is very close, with the semantic feedback-based scene recognition achieving 94.25% accuracy, a slight improvement of 0.25% compared to the traditional method's 94% accuracy. Simulation results demonstrate that when performing scene recognition, the channel feature statistics obtained from the semantic database can achieve accuracy similar to accurate channel features, requiring only a small amount of semantic database labels as feedback.

[0156] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for UAV air-to-ground information sharing and integrated communication based on semantic communication, characterized in that: The method includes the following steps: Step S1: Construct a 3D model of the UAV-assisted ISAC system; The 3D model of the ISAC system includes a ground station, a drone, and the communication link between the ground station and the drone. Step S2: Analyze the CSI in the three-dimensional model of the ISAC system to obtain channel characteristic information in the time domain, spatial domain, and Doppler domain; Step S3: Based on the acquired CSI characteristic information, formulate a CSI feedback plan. The specific operations are as follows: Step S3-1: Construct a CSI semantic database based on deep clustering, including the following steps: Step S3-1-1: Use channel samples CSI generates a semantic library, in which, It is the first CIR of each channel, It is the number of samples; Step S3-1-2: Generate the channel subspace using the AE algorithm and the K-means algorithm. The AE algorithm includes an encoder and a decoder; the encoder compresses the input channel data into low-dimensional features. ； in, These are the encoded features. and These are the encoder's weight matrix and bias vector; It is the activation function; the AE loss in the network is represented as: ； The decoder will compress the features As input, and reconstruct data that is as similar as possible to the original data: ； in, It's about reconstructing the data. and These are the weight matrix and bias vector of the decoder; Step S3-1-3: Cluster the low-dimensional features obtained by the encoder using K-means; Step S3-2: Identify the CSI semantic database. The specific operation is as follows: Step S3-2-1: Transfer new channel data The input is fed into a trained autoencoder to produce two features, and the distance between the new channel data and all K current cluster centers is calculated. Step S3-2-2: Assign the new channel data to the cluster represented by the nearest cluster center, and update the learning rate parameter of the selected cluster center. The learning rate parameter is: ； in, It is the old cluster centroid. It is the new cluster centroid; The range (0,1) is used to control the magnitude of cluster centroid updates; Step S3-2-3: Set the distance and threshold between the new and old clusters. The comparison is performed; if the result is less than the threshold, the new channel data belongs to the cluster; if the result is greater than the threshold, it means that a new channel subspace structure has emerged, and all channel data need to be re-clustered. Step S3-3: Construct and generate a feedback scheme based on the CSI semantic database and recognition results; Step S4: Verify the established CSI feedback method.

2. The UAV air-to-ground integrated sensing communication method based on semantic communication according to claim 1, characterized in that: The communication link between the UAV and the ground station is affected by distance, obstacles, and the speed of the UAV, and these effects are reflected by the Channel Identity System (CSI). The CSI describes the channel's behavior in the time delay domain using the Channel Indicator (CIR), expressed as follows: ； in, The complex amplitude for each path, Delay for each path, Horizontal angle of arrival for each path Vertical angle of arrival for each path and The Doppler frequency shift is for each path, and L is the total number of multipaths; Let be the impulse function; It is the imaginary unit.

3. The UAV air-to-ground integrated sensing communication method based on semantic communication according to claim 1, characterized in that: In step S3-1-3, K-means is used to cluster the low-dimensional features obtained by the encoder. The specific operation is as follows: First, randomly select k data points from the channel dataset as initial centroids, and calculate the distance between each data point and all centroids; then assign each data point to the nearest centroid to form k clusters. For each cluster, a new centroid is calculated by averaging all data points assigned to that cluster, and the updated centroid is compared with the previous centroid. If the centroid does not change significantly, or the maximum number of iterations has been reached, stop the algorithm; otherwise, repeat the process. Once the algorithm converges, it returns the final centroids and cluster assignments; the core metric is the sum of squared errors (SSE). ； in, It is a semantic database The center.

4. The UAV air-to-ground integrated sensing communication method based on semantic communication according to claim 1, characterized in that: The AE algorithm is used for feature learning of channel data, and the K-means function is used for clustering features. The clustering loss of the K-means function is expressed as: ； in, This indicates whether a channel sample belongs to the k-th semantic database; 1 indicates belonging and 0 indicates not belonging. By updating the loss function, joint training of feature learning and clustering processes is achieved: ； in, and It is a weighting factor.

5. The UAV air-to-ground integrated sensing communication method based on semantic communication according to claim 1, characterized in that: Step S4 involves validating the established CSI feedback method, including using profile coefficients. and The measures are used to evaluate clustering performance and to validate wave velocity prediction and scene recognition in the development of CSI feedback schemes; The contour coefficient Used to compare the performance of different clustering algorithms, it depends only on the actual distribution of the clusters; silhouette coefficient The definition is as follows: ； in, express The average distance between the MPC and other MPCs belonging to the same cluster. express The minimum distance between the MPC and the MPC belonging to a different cluster; Within the range of [-1, 1], a larger value indicates better clustering performance; The A measure is a robust metric used to balance precision and recall; The measure is defined as follows: ； in, and This represents the average of precision and recall. Within the range of [−1,1], a larger value indicates better clustering robustness.

Images