Intelligent online construction method and device for spectrum map of low-altitude networked unmanned aerial vehicle

Abstract

The invention discloses an intelligent online construction method and device for a spectrum map of a low-altitude networked unmanned aerial vehicle. The intelligent online construction method comprises the following steps: step 1, dividing a model at a low altitude; step 2, constructing a frequency spectrum prediction model; and step 3, an intelligent frequency spectrum prediction method based on online learning. According to the method, a corresponding online construction framework is provided according to the real-time spectrum map prediction requirement of the future trajectory of the low-altitude networked unmanned aerial vehicle, and in particular, full-amount sampling, random sampling and near sampling strategies are respectively designed according to different spectrum characteristics and networking requirements, so that the real-time spectrum map prediction requirement of the future trajectory of the low-altitude networked unmanned aerial vehicle is met. The current network verification result shows the effectiveness of the proposed method and the reliability of the designed strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a low-altitude network-connected unmanned aerial vehicle spectrum map intelligent online construction method and device. BACKGROUND

[0002] In the current field of wireless communication, with the increasing saturation of ground networks and the continuous opening of low-altitude airspace, the deployment and optimization of low-altitude network-connected (UAV) communication networks are gradually becoming a hot spot of research and application. Such networks not only significantly enhance the extension capability of ground networks in the air, at a distance, or in special scenarios, but also widely serve key application scenarios such as emergency communication, environmental monitoring, and logistics distribution. However, a core challenge faced by low-altitude communication networks is how to achieve efficient management and intelligent scheduling of spectrum resources in a dynamic and complex electromagnetic environment, with the construction and real-time updating of spectrum maps being particularly critical. If unmanned aerial vehicles can perceive the current spectrum usage situation and predict future spectrum occupation, they can dynamically adjust communication parameters, significantly improve communication quality, and effectively avoid interference.

[0003] Current spectrum map construction methods still have many limitations. Traditional construction methods based on empirical statistics, although low in cost, are difficult to adapt to the characteristics of rapid spectrum changes due to their reliance on historical average data, and lack timeliness and adaptability. Although the random geometry or ray tracing methods based on physical modeling have more theoretical depth, they rely on accurate environment modeling and have high computational complexity, which restricts their practicality. In recent years, deep learning methods have been introduced into spectrum prediction tasks, which can obtain high-precision prediction results with the support of big data, but they are still mostly offline training, which is difficult to meet the high real-time demand of "learning while flying" in low-altitude communication scenarios. At the same time, current spectrum map construction methods have obvious limitations in processing spatio-temporal features. Existing deep learning models mostly use CNN or RNN, which are difficult to accurately capture key spatio-temporal correlations. In low-altitude scenarios, spectrum is affected by spatial terrain and time dimension signal path, and the equalization processing of traditional models easily masks key information such as sudden interference, reducing accuracy. The nonlinear spatio-temporal coupling characteristics of spectrum in complex environments are difficult to model, the existing models have limited receptive fields, are difficult to handle long-range dependencies, and are prone to cumulative errors. Moreover, the spatio-temporal distribution of low-altitude spectrum evolves dynamically, and the existing models lack dynamic weight adjustment capability, respond lag to spectrum mutations, and are difficult to meet the needs of high-dynamic scenarios.

[0004] In some deep learning recommendation systems, online learning techniques are used to practice full-sample training (training a model with full data) and random sampling training (random sampling all samples without distinction); in the graph neural network calculation module, there are also related contents of random sampling and neighborhood sampling, which are used to solve the neighborhood explosion problem and storage bottleneck of large-scale graph calculation. However, none of these directly addresses the real-time spectrum prediction problem. SUMMARY

[0005] In view of the challenges brought by dynamic changes of spectrum resources and high maneuverability of unmanned aerial vehicles in a low-altitude network environment, the present application proposes a low-altitude networked unmanned aerial vehicle spectrum map intelligent online construction method and device that fuses a space-time attention mechanism and an adaptive sampling strategy, realizes high-precision real-time prediction and dynamic updating of the spectrum map through multi-dimensional feature fusion, dynamic model parameter updating and a differential historical sample screening mechanism, and significantly improves the spectrum resource management capability and system adaptability of the low-altitude unmanned aerial vehicle network in a complex electromagnetic environment.

[0006] To achieve the above technical purposes, the technical solution adopted by the present application is as follows:

[0007] In a first aspect, the present application discloses a low-altitude networked unmanned aerial vehicle spectrum map intelligent online construction method, which comprises the following steps:

[0008] S1, according to the size of the target low-altitude area, the target low-altitude area is divided into multiple cubes, and each cube and its absolute position are defined;

[0009] S2, modeling the base stations and service frequency bands involved in the target low-altitude area, defining the total received signal strength of the cube level; making the networked unmanned aerial vehicle fly along a preset trajectory, representing the flight path of the unmanned aerial vehicle as a position sequence containing a time stamp, and periodically measuring the total signal reception strength of the cube where the unmanned aerial vehicle is located during flight, and obtaining a certain number of sampling spectrums in each time interval, and constructing sample data, each group of sample data including the sampling time stamp, position coordinates, wireless signal quality data and total signal reception strength of the cube corresponding to the sampling spectrum;

[0010] S3, constructing a space-time attention prediction model, which extracts features from the input sampling time stamp, position coordinates, wireless signal quality data and sampling spectrum to obtain time features, space features, wireless signal quality features and network frequency band features respectively, and splices to obtain an initial input feature matrix; then encodes the position sequence containing the time stamp into a position vector to generate a position encoding matrix, and adds the generated position encoding matrix to the initial input feature matrix to obtain an input feature matrix; then inputs the input feature matrix into the space-time attention layer for fusion to capture the signal quality changes at different times in the same space, the signal distribution at the same time in adjacent spaces, and the signal differences at different frequencies in the same space; finally, the fusion features output by the space-time attention layer are subjected to nonlinear transformation to strengthen the nonlinear association between space, signal and frequency band, and the nonlinear transformation result is output as a wireless signal reception strength prediction value;

[0011] S4, according to different spectrum characteristics and network demand, respectively proposes online learning strategies based on full sampling training, random sampling training and adjacent sampling training, and uses the constructed spatiotemporal attention prediction model to perform real-time spectrum prediction.

[0012] Step S1 further includes:

[0013] The target low-altitude area is subdivided into N cubes, and the size of the target low-altitude area is UxVxW m 3 , respectively corresponding to the longitude, latitude and height range, N=N U xN W xN V , N U , N W and N V represent the discretization number along the longitude, latitude and height axes, respectively;

[0014] The cube set is defined as R={1,…,r n ,…,r N}, and the absolute position matrix X (R) of all cubes is represented as wherein r n represents the absolute position vector of the cube r K , and and are the longitude, latitude and height, respectively.

[0015] Further, in step S2, the target low-altitude area involved in the base station and the service frequency band is modeled, and the process of defining the full amount of received signal strength at the cube level includes:

[0016] The target low-altitude area involved in the base station and the service frequency band is modeled, and the base station set is defined as wherein b K represents the kth base station; each base station b K has a fixed geographic location and frequency band; for all base stations, the frequency band set involved is wherein f M represents the mth frequency band;

[0017] For any cube r n , the signal reception strength of each base station b K on the frequency band f M is represented as RSSI n,m,k , and the total signal strength of multiple base stations on the frequency band f M in the cube is:

[0018]

[0019] Wherein, n represents the number of low-altitude areas discretized along the longitude, latitude and height axis, m represents the index of the frequency band, and k represents the index of the base station.

[0020] Further, in step S3, the space-time attention model comprises a feature extraction layer, a feature embedding layer, a position encoding layer, a space-time attention layer and a fully connected layer.

[0021] The feature extraction layer extracts features from the input sampling timestamp, position coordinates, wireless signal quality data and sampling spectrum, respectively obtaining time features, space features, wireless signal quality features and network frequency band features; wherein the wireless signal quality data includes synchronization signal received strength, signal-to-noise ratio, signal reception quality and received power.

[0022] The feature embedding layer maps the input feature vector including time features, space features, wireless signal quality features and network frequency band features to a high-dimensional feature space through linear transformation to obtain an initial input feature matrix X embed ;

[0023] The position encoding layer encodes the position sequence containing the time stamp into a position vector to generate a position encoding matrix PE, and adds the generated position encoding matrix PE and the initial input feature matrix X embed to obtain an input feature matrix X embed + PE.

[0024] The space-time attention layer maps the input feature matrix to multiple subspaces and calculates the attention function in parallel, fuses the time features, space features, wireless signal quality features and network frequency band features; wherein the multiple subspaces learn the dependency relationship between features through attention mechanism, capture the potential law of different feature combinations through parallel calculation, and then fuse to form a context representation containing multi-feature interaction information, realizing comprehensive extraction of feature dependency relationship in sequence samples.

[0025] The fully connected layer is composed of multiple linear layers and activation functions, which performs nonlinear transformation on the fused features output by the space-time attention layer to strengthen the nonlinear association of space-time, signal and frequency band, and finally outputs the wireless signal received strength prediction value of each sample.

[0026] Further, the position encoding layer indirectly embodies the time sequence information through sequence indexes, and generates a position encoding matrix of a fixed dimension based on a sine-cosine function of the sequence indexes: for each sample in the sequence, a frequency factor is calculated through the model dimension, a coding value is generated by using a sine function for an even dimension, and a coding value is generated by using a cosine function for an odd dimension, thereby forming a position encoding matrix with a dimension of [maximum sequence length, model dimension], and a buffer fixed parameter is used to avoid gradient update; wherein, the position encoding layer uses a feature fusion mechanism to process the forward propagation process, specifically, in the forward propagation process, the part of the position encoding matrix matched with the input sequence length is directly superimposed with the embedded feature vector, so that the output feature contains both the original input feature and the time sequence position association information of the sample in the sequence.

[0027] Further, the spatio-temporal attention layer captures the dependency relationship between sequence samples through a multi-head attention mechanism, including a feature embedding module, an attention calculation module, a residual normalization module, and an output mapping module.

[0028] The feature embedding module maps the input feature to the model dimension by using linear transformation, the linear transformation matrix maps the input feature (including grid coordinates, frequency band information, time stamp, and signal quality indicators, etc.) to a unified model dimension, thereby realizing preliminary fusion of multiple features through dimension unification, and providing a consistent feature space for subsequent attention calculation;

[0029] The attention calculation module based on the multi-head attention mechanism simultaneously inputs the feature vector after position encoding as query, key, and value, and extracts feature dependency relationship from different subspaces through multiple attention heads in parallel; wherein, the projection matrix of the query, key, and value of the multi-head attention is learned adaptively by the model through training, so as to capture the association relationship between the input features (including grid coordinates, frequency band information, time stamp, and signal quality indicators, etc.); the original dimensions of the query matrix, the key matrix, and the value matrix are constructed based on the unified feature dimension after fusion, and the features are accurately mapped in the subspaces through the product with the projection matrix, the attention calculation captures the association between the spatial distribution of the grid coordinates, the network characteristics of the frequency band information, and the implicit features of the time and signal quality, so that the feature association of the frequency spectrum situation sample forms the learned association relationship with the core data features inputted;

[0030] The residual normalization module enhances the training stability by performing layer normalization processing on the residual connection between the attention output and the original input feature;

[0031] The output mapping module adopts a full connection layer sequence, first expands the feature dimension, and is processed by an activation function and regularization, and finally is mapped to a target output dimension to fuse the associated features of the sequence samples; specifically, the full connection layer is composed of multiple linear layers and activation functions, and performs nonlinear transformation on the features fused with the four types of collected features processed by the spatio-temporal attention layer, strengthens the nonlinear association between space, signal and frequency band, and finally outputs the wireless signal RSSI prediction value of each sample.

[0032] Further, in step S4, according to different spectrum characteristics and network connection requirements, online learning strategies based on full sampling training, random sampling training and adjacent sampling training are respectively proposed, and the process of real-time spectrum prediction by using the constructed spatio-temporal attention prediction model includes:

[0033] According to the signal attenuation and Doppler effect spectrum characteristics in wireless communication, a preset rate threshold for judging whether the spectrum is in a stable state is set, and a preset sample threshold is determined in combination with the time correlation of the spectrum characteristics, the preset sample threshold is used to ensure that the samples can fully reflect the distribution law of the stable spectrum; a communication delay threshold for representing the prediction real-time requirement is determined based on the communication delay requirement of the multi-machine cooperative operation of the unmanned aerial vehicle;

[0034] For the case that the spectrum dynamic change rate is lower than the preset rate threshold and the total number of samples is less than the preset sample threshold, all samples are selected from the new experience pool every d batches for training, and all available historical samples are selected for model training during training;

[0035] For the case that the spectrum dynamic change rate is lower than the preset rate threshold, the total number of samples is greater than the preset sample threshold, and the communication delay is less than the communication delay threshold, d2 samples are randomly sampled from the experience pool every d batches for training;

[0036] For the case that the spectrum dynamic change rate is higher than the preset rate threshold, the total number of samples is greater than the preset sample threshold, and the communication delay is less than the communication delay threshold, the most recent d2 samples are sampled from the experience pool every d batches for training.

[0037] In a second aspect, the application discloses a low-altitude network-connected unmanned aerial vehicle spectrum map intelligent online construction device based on the foregoing method, characterized in that the device comprises:

[0038] A cubic division and definition module is configured to divide a target low-altitude area into a plurality of cubes according to the size of the target low-altitude area, and define each cube and its absolute position.

[0039] A spectrum modeling module is configured to model a target low-altitude area involving a base station and a service frequency band, define a full amount of received signal strength at a cubic level, and make a network-connected unmanned aerial vehicle (UAV) fly along a preset trajectory and periodically measure a total signal strength of a cubic where the UAV is located during flight, so that a certain number of sampled spectrums in a corresponding time interval are obtained in each time slot.

[0040] A real-time spectrum prediction module is configured to construct a spatio-temporal attention prediction model, the input data of the spatio-temporal attention prediction model being grid coordinates of the UAV and frequency band information, and the output being a predicted spectrum, and the real-time spectrum prediction module is configured to propose online learning strategies based on full amount sampling training, random sampling training and nearby sampling training according to different spectrum characteristics and network connection requirements, and use the constructed spatio-temporal attention prediction model to perform real-time spectrum prediction.

[0041] Compared with the prior art, the low-altitude network-connected UAV spectrum map intelligent online construction method and device have the following beneficial effects:

[0042] The low-altitude network-connected UAV spectrum map intelligent online construction method and device use deep learning technology, are suitable for complex scenarios caused by dynamic changes of spectrum resources and high maneuverability of the UAV in a low-altitude network-connected environment, can adapt to different spectrum characteristics and network connection requirements by optimizing a spatio-temporal attention model architecture design and a sample selection strategy, realize high-precision real-time prediction and dynamic updating of a spectrum map, and significantly improve spectrum resource management capability and system adaptability of a low-altitude UAV network in a complex electromagnetic environment. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flowchart of a low-altitude network-connected UAV spectrum map intelligent online construction method is shown in the drawings.

[0044] Figure 2 A low-altitude route measurement trajectory is shown in the drawings.

[0045] Figure 3 A performance graph of three strategies in a first case is shown in the drawings.

[0046] Figure 4 A performance graph of three strategies in a second case is shown in the drawings.

[0047] Figure 5 A comparison diagram of predicted values and real values of future samples in different strategies is shown in the drawings. DETAILED DESCRIPTION

[0048] Embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0049] Figure 1For the flow chart of the low-altitude networked unmanned aerial vehicle spectrum map intelligent online construction method involved in the embodiment of the application, the low-altitude networked unmanned aerial vehicle spectrum map intelligent online construction method of the application specifically comprises the following steps:

[0050] 1. Low-altitude division model

[0051] According to the size of the target low-altitude area, the target low-altitude area is divided into a plurality of cubes, and each cube and its position are defined.

[0052] Consider a low-altitude area with a size of UxVxW m 3 , which respectively correspond to the longitude, latitude and height range of the low-altitude area. The area is subdivided into N cubes, where N=N U xN W xN V , N U , N W and N V represent the discretization number along the longitude, latitude and height axes, respectively, and the cube set can be defined as R={1,…,r n ,…,r N}. For cube r n , the absolute position vector can be defined as where and are the longitude, latitude and height, respectively. Therefore, the absolute position matrix X (R) of all cubes is defined as

[0053] 2. Construction of spectrum prediction model; further, the base stations and service frequency bands involved in the target low-altitude area are modeled, so as to define the full amount of Received Signal Strength Indicator (RSSI) at the cube level. Considering that the networked unmanned aerial vehicle flies along the preset trajectory, the real-time spectrum prediction problem is established, and the spatiotemporal attention prediction model is constructed to predict the spectrum;

[0054] There are a plurality of base stations in the area, and the base station set is where b K represents the kth base station. Each base station b K has a fixed geographic position and frequency band. For all base stations, the frequency band set involved is where f M represents the mth frequency band. For any cube r n , the signal received strength (Received Signal Strength Indicator, RSSI) of each base station b K on the frequency band f M can be represented asn,m,k The total signal strength of the plurality of base stations in the cube on the frequency band f M is:

[0055]

[0056] where n represents the number of discretization along the longitude, latitude and height axis of the low-altitude area, m represents the index of the frequency band, and k represents the index of the base station.

[0057] Consider that a network-connected unmanned aerial vehicle flies along a predetermined trajectory in a low-altitude area. The unmanned aerial vehicle will periodically measure the RSSI n,m of its current cube during flight. The flight path and position of the unmanned aerial vehicle can be represented as a time series A certain number of sampled RSSIs in this time interval can be obtained for each time slot t. In order to ensure the reliability of the network connection and avoid spectral holes, the spectrum, i.e. RSSI, of the future trajectory needs to be predicted.

[0058] The obtained data features include four types of core information: first, the time feature, i.e. the time stamp recorded by the PC end and the user equipment, used to mark the time sequence of data collection and the time synchronization between devices; second, the spatial feature, high-precision longitude and latitude coordinates, accurately positioning the geographical location of the collection point; third, the wireless signal quality feature, including the received strength of the synchronization signal, the signal-to-interference-plus-noise ratio, the reception quality, the received power, etc., directly reflecting the transmission performance of the 5G signal; and fourth, the network frequency band feature, identifying the currently connected 5G frequency band, correlating the coverage range and transmission rate of the signal. These data support the intelligent online construction of the low-altitude network-connected unmanned aerial vehicle spectrum map, analyze the signal distribution through signal features and spatio-temporal information, optimize the unmanned aerial vehicle communication coverage, dynamically adapt to signal fluctuations, ensure multi-machine collaborative operation, and solve the communication problems in actual deployment.

[0059] The spatio-temporal attention model mainly consists of a feature embedding layer, a position encoding layer, a spatio-temporal attention layer, and a fully connected layer.

[0060] The feature embedding layer maps the input feature vector including the grid coordinates, frequency band information, time stamp, and wireless signal quality feature to a high-dimensional feature space through linear transformation, obtaining an initial input feature matrix X embed . The feature dimension is assigned a weight, the grid coordinates and time stamp weight are strengthened for spatio-temporal positioning, the frequency band and signal quality weight highlight the signal performance correlation, and the matrix source fuses the core features collected.

[0061] The core goal of the position encoding layer is to provide the model with the relative or absolute position information of the elements in the sequence. When processing spatio-temporal data, this is particularly important for capturing the sequential characteristics of the signal in time and space. Position encoding encodes the position information into a vector, which is then added to the input features, so that the model can learn the position-related patterns.

[0062] The position encoding layer generates a fixed-dimension position encoding matrix based on the sine-cosine function: for the position-th sample in the sequence, a frequency factor is calculated through the model dimension, and for the even-numbered dimensions, the encoding value is generated by using the sine function (position multiplied by the sine value of the frequency factor), and for the odd-numbered dimensions, the encoding value is generated by using the cosine function (position multiplied by the cosine value of the frequency factor), forming an encoding matrix with a dimension of [maximum sequence length, model dimension], and a buffer fixed parameter is used to avoid gradient update; the position encoding layer uses a feature fusion mechanism to process the normal propagation process, specifically, in the forward propagation process, the part of the encoding matrix matched with the input sequence length is directly superimposed with the embedded feature vector, so that the output feature contains both the original input feature and the time sequence position information of the sample in the sequence.

[0063] In the present application, the latitude and longitude are first converted into grid coordinates through rasterization processing, and the original timestamp is retained as time sequence information, so that the processed samples contain both spatial features (grid coordinates) and time features (timestamps); during data processing, core features such as network frequency band and signal quality are retained, forming a multi-dimensional feature matrix containing space-time and network characteristics; the processed feature matrix covers the spatial dimension of grid coordinates, the time sequence dimension of timestamps, and the identification dimension of frequency band information, laying a data foundation for subsequent model learning of the association between these features.

[0064] The position encoding layer encodes the position information containing the timestamp time sequence and the grid coordinates into a vector, and generates a position encoding matrix PE and an initial input feature matrix X embed are added to obtain the input feature matrix X of the space-time attention layer X = X embed + PE. The encoding vector quantifies the timestamp sequence and the grid coordinate distance, and after superposition, it strengthens the binding of the space-time features and the signal and frequency band features, and conforms to the collection logic.

[0065] The spatio-temporal attention layer adopts a multi-head attention mechanism, which can capture the dependency relationship between samples and fuse the context representation of all sample information. The spatio-temporal attention layer includes a feature embedding module, an attention calculation module, a residual normalization module and an output mapping module; the feature embedding module maps the input feature to the model dimension by linear transformation; the attention calculation module inputs the feature vector after position encoding as query, key and value based on the multi-head attention mechanism, extracts feature dependency relationship from different subspaces through parallel multiple attention heads; the residual normalization module connects the attention output and the original input feature in residual, and then performs layer normalization processing to enhance the training stability; the output mapping module adopts a full connection layer sequence, first expands the feature dimension and then performs activation function and regularization processing, and finally maps to the target output dimension to fuse the associated features of the sequence samples. The multi-head attention mechanism captures information from different subspaces through parallel calculation of multiple attention heads, thereby improving the expression ability of the model. In the multi-head attention calculation process of the present application, for spectral data, the subnetwork accurately calculates the query matrix, key matrix and value matrix, and uses the attention mechanism to closely associate the features of different subsequence positions of the spectrum, thereby effectively capturing the complex long-range dependency relationship of the spectral situation sample. The multi-head attention mechanism maps the input to multiple different subspaces and calculates multiple attention functions in parallel, so that the model can capture different types of dependency relationships. The outputs of each head are spliced or averaged. In the present application, the number of attention heads is set as a hyperparameter to adapt to the learning needs of the model for feature interaction relationship; the output part processes the features after attention fusion through a full connection layer sequence, and finally maps to the target output dimension to realize comprehensive fusion of multiple features. The original dimensions of the query matrix, the key matrix and the value matrix are constructed based on the uniform dimension of the input feature after mapping by the embedding layer, and the subspace mapping is completed through the product of the projection matrix built-in the multi-head attention; the attention calculation learns to capture the spatial distribution of grid coordinates, the network characteristics of frequency band information, and the association between time and signal quality hidden features, so that the feature association of the spectral situation sample forms an associated relationship with the input core data features through model autonomous learning. The full connection layer is composed of multiple linear layers and activation functions, which performs nonlinear transformation on the features fused with four types of collected features after processing by the spatio-temporal attention layer, strengthens the nonlinear association of space-time-signal-frequency, and finally outputs the wireless signal RSSI prediction value of each sample.

[0066] 3. Different online learning sampling methods are proposed for different spectral characteristics;

[0067] Since the RSSI samples of this scenario are not fixed but constantly acquired, an online learning method is adopted to solve this problem. Specifically, by dynamically receiving new samples and using the new sample experience pool for training and real-time inference, the real-time and accuracy of future spectrum prediction are ensured. For example, the learning model adopts a spatio-temporal attention network, which inputs the location and frequency band of the sample and outputs the corresponding RSSI. The historical spectrum samples are used to train the neural network, so that it can obtain the ability to predict the future trajectory RSSI.

[0068] During the training process of the learning model, every d epochs, the historical RSSI samples collected in the current experience pool are used to predict the signal strength of the next d1 sampling points. After each training of d epochs, the experience pool is updated to add the last d1 new samples to maintain the real-time and adaptability of the model. According to the different characteristics of the spectrum situation, the present application proposes the following three strategies:

[0069] Strategy 1: For the case where the spectrum change rate is lower than the preset rate threshold and the total sample number is less than the preset sample threshold, all samples are selected from the new experience pool for training every d batches, and all available historical samples are selected for model training during training. The preset rate threshold is set according to the signal attenuation, Doppler effect and other spectrum characteristics in wireless communication to determine whether the spectrum is in a stable state; the preset sample threshold is determined in combination with the time correlation of the spectrum characteristics to ensure that the samples can fully reflect the distribution law of the stable spectrum;

[0070] Strategy 2: For the case where the spectrum dynamic change is lower than the preset rate threshold and the total sample number is greater than the preset sample threshold, and the prediction real-time requirement is high, d2 samples are randomly sampled from the experience pool for training every d batches. The requirement for prediction real-time is determined based on the communication delay requirement of multi-robot cooperative operation of unmanned aerial vehicles, and the value of d2 is set to avoid the influence of redundant historical data on real-time and ensure the representativeness of samples on stable spectrum distribution;

[0071] Strategy 3: For the case where the spectrum dynamic change is higher than the preset rate threshold, the total sample number is greater than the preset sample threshold, or the real-time requirement is higher than the preset time threshold, the last d2 samples are sampled from the experience pool for training every d batches. The preset rate threshold of the spectrum dynamic change corresponds to the signal mutation scenario caused by the rapid movement of unmanned aerial vehicles, and the preset time threshold of the real-time requirement is related to the low delay requirement of unmanned aerial vehicle formation control, etc. The latest signal characteristics under the rapid change of the spectrum are captured.

[0072] The classification logic of the three strategies takes the core elements of the spectrum prediction scene as the anchor point, and the threshold setting is based on the wireless communication physical characteristics and the actual needs of the unmanned aerial vehicle operation, which is significantly different from the sampling strategies in the prior art that are designed only around a single general dimension, and realizes deep adaptation to the real-time spectrum prediction scene.

[0073] The simulation analysis is as follows:

[0074] In order to verify the performance of the proposed method, a network terminal carried by a UAV is used for low-altitude drive testing in Guanbeicun, Nanchang, Jiangxi. The flight height of the UAV is 150 meters, and every 20 meters x 20 meters is a cube size in the horizontal and vertical directions. The trajectory of the visual low-altitude drive test is shown in Figure 2 Further exemplarily, let the first group d, d1, d2 be 100, 200, 800 respectively, and the second group d, d1, d2 be 100, 400, 800 respectively. The total epochs are set to 1500, and the prediction spectrum prediction root mean square error (RMSE) performance is evaluated once every 100 epochs on the future 400 samples. The experimental results of the three strategies will be described next.

[0075] Figure 3 The validation RMSE performance of the proposed three online learning strategies in the first group d, d1, d2 is shown in the test samples corresponding to each epoch. As can be seen from the figure, on average, the average RMSE of the three strategies is 9.4166dB, 10.9559dB, and 9.6713dB respectively.

[0076] Figure 4 The validation RMSE performance of the proposed three online learning strategies in the second group d, d1, d2 is shown in the test samples corresponding to each epoch. As can be seen from the figure, on average, the average RMSE of the three strategies is 9.3795dB, 9.7191dB, and 9.5498dB respectively. The strategy using the full amount of samples can be regarded as the baseline performance, and the other two strategies are not far from the baseline performance, which to some extent illustrates the reliability of strategies two and three in the proposed online learning method. Different strategies can be selected according to different needs. In addition, the curves of strategy 1 and strategy 3 are very close, indicating that the learning performance of the model is mainly affected by the recent samples.

[0077] Figure 5 The prediction and true value comparison of different strategies on the trajectory of future samples at the 900th epoch and the 1500th epoch is shown. The prediction and error visualization of the three strategies are relatively small, which further illustrates the reliability of the proposed method.

[0078] While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the present disclosure without departing from the spirit and scope of the application. Accordingly, it is intended that such additions and modifications be included within the scope of the application. It is the following claims, including any amendments thereto, which define the scope of the application.

[0079] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for intelligent online construction of spectrum maps for low-altitude networked drones, characterized in that: The method comprises the following steps: S1, divide the target low-altitude area into multiple cubes according to its size, and define each cube and its absolute position; S2: Model the base stations and service frequency bands involved in the target low-altitude area and define the total received signal strength at the cube level. The connected drone is instructed to fly along a preset trajectory, representing the drone's flight path as a time-stamped position sequence. The total received signal strength of the cube in which it is located is regularly measured during flight. A certain number of sample spectra are obtained within the corresponding time interval in each time slot to construct sample data. Each set of sample data includes the sampling timestamp, location coordinates, wireless signal quality data, and the total received signal strength of the cube in which it is located. S3, constructing a spatiotemporal attention prediction model, which extracts features from the input sampling timestamp, location coordinates, wireless signal quality data, and sampling spectrum to obtain time features, spatial features, wireless signal quality features, and network frequency band features, respectively, and splices them to obtain an initial input feature matrix; then encoding the position sequence containing the timestamp into a position vector to generate a position coding matrix, and adding the generated position coding matrix to the initial input feature matrix to obtain an input feature matrix; then inputting the input feature matrix into the spatiotemporal attention layer for fusion to capture the signal quality changes in different time series of the same space, the signal distribution in adjacent spaces of the same sequence, and the signal differences in different frequency bands of the same space and time; finally, performing a nonlinear transformation on the fused features output by the spatiotemporal attention layer to strengthen the nonlinear correlation between time and space, signals, and frequency bands, and outputting a predicted value of the wireless signal reception strength based on the nonlinear transformation result; S4, according to different spectrum characteristics and network requirements, proposes online learning strategies based on full sampling training, random sampling training, and adjacent sampling training, and uses the constructed spatiotemporal attention prediction model to perform real-time spectrum prediction.

2. The method for intelligent online construction of spectrum map of low-altitude networked UAVs according to claim 1 is characterized in that: Step S1 further comprises: The target low-altitude area is subdivided into N cubes, and the size of the target low-altitude area is U×V×W m 3 , corresponding to the longitude, latitude and altitude ranges respectively, N = N U ×N W ×N V , N U 、N W and N V represent the discretization quantities along the longitude, latitude, and altitude axes, respectively; The cube set is defined as R = {1,…,r n ,…,r N }, the absolute position matrix X of all cubes (R) Expressed as Represents a cube r n The absolute position vector, and They are longitude, latitude and altitude respectively.

3. The method for intelligent online construction of spectrum map of low-altitude networked UAVs according to claim 1 is characterized in that: In step S2, the process of modeling the base stations and service frequency bands involved in the target low-altitude area and defining the total received signal strength at the cube level includes: Model the base stations and service frequency bands involved in the target low-altitude area, and let the base station set be where b K represents the kth base station; each base station b K There are fixed geographical locations and frequency bands; for all base stations, the set of frequency bands involved is where f M represents the mth frequency band; For any cube r n , each base station b K In the frequency band f M The signal receiving strength is expressed as RSSI n,m,k , then the frequency band f of multiple base stations in this cube M The total signal strength on is: Among them, n represents the number of low-altitude areas discretized along the longitude, latitude and altitude axes, m represents the index of the frequency band, and k represents the index of the base station.

4. The method for intelligent online construction of spectrum maps of low-altitude networked drones according to claim 1 is characterized in that: In step S3, the spatiotemporal attention model includes a feature extraction layer, a feature embedding layer, a position encoding layer, a spatiotemporal attention layer, and a fully connected layer; The feature extraction layer extracts features from the input sampling timestamp, location coordinates, wireless signal quality data, and sampled spectrum to obtain temporal features, spatial features, wireless signal quality features, and network frequency band features, respectively. The wireless signal quality data includes synchronization signal reception strength, signal-to-noise ratio, signal reception quality, and reception power. The feature embedding layer maps the input feature vectors including time features, spatial features, wireless signal quality features and network frequency band features to a high-dimensional feature space through linear transformation to obtain the initial input feature matrix X embed ; The position encoding layer encodes the position sequence containing the timestamp into a position vector to generate a position encoding matrix PE, and compares the generated position encoding matrix PE with the initial input feature matrix X embed After adding, we get the input feature matrix X = X embed +PE; The spatiotemporal attention layer maps the input feature matrix to multiple subspaces and parallelly computes attention functions to fuse temporal features, spatial features, wireless signal quality features, and network frequency band features. Multiple subspaces autonomously learn the dependencies between features through the attention mechanism, capture the underlying patterns of different feature combinations through parallel computation, and then fuse them to form a contextual representation containing multi-feature interaction information, achieving comprehensive extraction of feature dependencies in sequence samples. The fully connected layer consists of multiple linear layers and activation functions, which performs nonlinear transformation on the fused features output by the spatiotemporal attention layer to strengthen the nonlinear correlation between time, space, signals and frequency bands, and finally outputs the predicted value of the wireless signal reception strength for each sample.

5. The method for intelligent online construction of spectrum map of low-altitude networked UAVs according to claim 4 is characterized in that: The position encoding layer indirectly reflects the temporal information through the sequence index, and generates a fixed-dimensional position encoding matrix based on the sine and cosine functions of the sequence index: for each sample in the sequence, the frequency factor is calculated through the model dimension, and the encoding value is generated by the sine function for even dimensions and the cosine function for odd dimensions, forming a position encoding matrix with the dimension [maximum sequence length, model dimension], and the parameters are fixed through the buffer to avoid gradient updates; wherein, the position encoding layer uses a feature fusion mechanism to process the forward propagation process. Specifically, during the forward propagation process, the part of the position encoding matrix that matches the input sequence length is directly superimposed with the embedded feature vector, so that the output feature contains both the original input feature and the temporal position association information of the sample in the sequence.

6. The method for intelligent online construction of spectrum maps of low-altitude networked drones according to claim 4 is characterized in that: The spatiotemporal attention layer captures the dependencies between sequence samples through a multi-head attention mechanism, including a feature embedding module, an attention calculation module, a residual normalization module, and an output mapping module; The feature embedding module uses linear transformation to map input features to model dimensions. The linear transformation matrix maps input features including grid coordinates, frequency band information, timestamps, and signal quality indicators to a unified model dimension, achieving preliminary fusion of multiple features through dimensional unification. The attention calculation module is based on a multi-head attention mechanism, which uses the position-encoded feature vector as query, key, and value input at the same time, and extracts feature dependencies from different subspaces through multiple parallel attention heads. The query, key, and value projection matrices of the multi-head attention are adaptively learned by the model through training to capture the correlation between input features. The original dimensions of the query matrix, key matrix, and value matrix are constructed based on the unified feature dimension after fusion, and the features are accurately mapped in the subspace by multiplication with the projection matrix. The attention calculation captures the spatial distribution of grid coordinates, the network characteristics of frequency band information, and the correlation between the implicit features of time and signal quality through learning, so that the feature correlation of the spectrum situation sample forms a learned correlation with the core data features of the input. The residual normalization module performs a residual connection between the attention output and the original input features, and then performs layer normalization to enhance training stability. The output mapping module uses a sequence of fully connected layers to first expand the feature dimension and process it through activation functions and regularization, and finally maps it to the target output dimension to fuse the associated features of the sequence samples. Specifically, the fully connected layer consists of multiple linear layers and activation functions, and performs nonlinear transformations on the features processed by the spatiotemporal attention layer and the fusion of four types of collected features, strengthening the nonlinear correlation between time, space, signal and frequency bands, and ultimately outputting the predicted RSSI value of the wireless signal for each sample.

7. The method for intelligent online construction of spectrum map of low-altitude networked UAVs according to claim 1 is characterized in that: In step S4, based on different spectrum characteristics and network requirements, online learning strategies based on full sampling training, random sampling training, and proximity sampling training are proposed. The process of using the constructed spatiotemporal attention prediction model to perform real-time spectrum prediction includes: Based on the signal attenuation and Doppler effect spectrum characteristics in wireless communications, a preset rate threshold is set to determine whether the spectrum is in a stable state. Combined with the time correlation of the spectrum characteristics, a preset sample threshold is determined to ensure that the sample can fully reflect the distribution pattern of the stable spectrum. Based on the communication delay requirements of multi-UAV collaborative operations, a communication delay threshold is determined to represent the real-time requirements of the prediction. When the rate of spectrum dynamic change is lower than the preset rate threshold and the total number of samples is less than the preset sample threshold, all samples are selected from the new experience pool for training every d batches, and all available historical samples are selected for model training during training; For the case where the spectrum dynamic change rate is lower than the preset rate threshold, the total number of samples is greater than the preset sample threshold, and the communication delay is less than the communication delay threshold, d2 samples are randomly sampled from the experience pool for training every d batches; For the case where the dynamic change rate of the spectrum is higher than the preset rate threshold, the total number of samples is greater than the preset sample threshold, and the communication delay is less than the communication delay threshold, the most recent d2 samples are sampled from the experience pool for training every d batches.

8. An intelligent online construction device for a spectrum map of a low-altitude networked UAV based on the method described in any one of claims 1 to 7, characterized in that: The device comprises: The cube division and definition module is used to divide the target low-altitude area into multiple cubes according to the size of the target low-altitude area, and define each cube and its absolute position; The spectrum modeling module is used to model the base stations and service frequency bands involved in the target low-altitude area and define the total received signal strength at the cube level. The connected drone is instructed to fly along a preset trajectory and periodically measures the total signal strength of the cube it is in during flight. A certain number of sampled spectra within the corresponding time interval are obtained in each time slot. The real-time spectrum prediction module is used to build a spatiotemporal attention prediction model. The input data of the spatiotemporal attention prediction model is the grid coordinates and frequency band information of the drone, and the output is the predicted spectrum. According to different spectrum characteristics and network connection requirements, online learning strategies based on full sampling training, random sampling training, and adjacent sampling training are proposed respectively, and real-time spectrum prediction is performed using the constructed spatiotemporal attention prediction model.

Images