Spectrum sensing method for three-dimensional broadband spectrum sensing communication system

Abstract

The invention relates to a spectrum sensing method for a three-dimensional broadband spectrum sensing communication system, and belongs to the field of wireless communication. The method comprises the steps that two types of input serve as information sources, an initial convolutional layer module extracts time domain and amplitude features, and a double-prior encoder module obtains a comprehensive prior vector; the decoupling convolution layer realizes space, frequency and time feature separation and is fused with prior in a feature linear modulation module; key channel characteristics are highlighted through a multi-head self-attention mechanism module, and a global pooling layer module outputs an occupancy probability and a normalized signal-to-noise ratio respectively; and through final joint loss function optimization, broadband spectrum occupancy detection and signal-to-noise ratio prediction in a three-dimensional air-ground environment are realized at the same time. According to the method, the problems that the detection performance is reduced and the generalization ability is insufficient in a complex channel environment, a long-distance low signal-to-noise ratio condition and dynamic spectrum change are solved.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a spectrum sensing method for a three-dimensional broadband spectrum sensing communication system. Background Technology

[0002] With the widespread deployment of drones in fields such as intelligent inspection, emergency communication, and logistics delivery, the demand for highly reliable broadband connections in low-altitude communication networks continues to grow. However, the air-to-ground channel environment has strong three-dimensional spatial characteristics and dynamism, and traditional fixed spectrum allocation methods can no longer meet the reliability and throughput requirements of drone networks. Therefore, dynamic spectrum access and spectrum sensing have become key technologies for improving spectrum utilization. Traditional spectrum sensing research has mainly focused on terrestrial cognitive radio networks (CRNs).

[0003] Typical techniques include energy detection (ED), cooperative detection, compressed detection, and deep learning-based broadband detectors. However, when spectrum sensing nodes are deployed on mobile drone platforms, the air-to-ground communication (ATG) channel alters the spatial and spectral characteristics of the received signal: the drone's altitude, trajectory, and attitude collectively contribute to path loss, shadow fading, and line-of-sight probability. Therefore, applying planar or static additive white Gaussian noise channel assumptions limits the generalization capability of existing spectrum sensing and detection models to 3D environments.

[0004] Current technologies employ dedicated air-to-ground channel models, such as geometry-based stochastic models (GBSM), measurement-driven fixed-wing UAV models, and multipath swarm surveys. However, no broadband sensing framework integrates 3D air-to-ground channel priors with deep learning detectors. GBSense and collaborative listening based on joint learning achieve robust performance at low signal-to-noise ratios (SNR), but still do not explicitly model UAV mobility or non-stationary air-to-ground channels. Furthermore, most existing technologies optimize for individual tasks (e.g., binary occupancy detection) while neglecting physical metrics such as SNR or power spectral density. In summary, existing technologies suffer from performance degradation and insufficient generalization capabilities in complex channel environments, long-distance low SNR conditions, and dynamic spectral changes. Summary of the Invention

[0005] This application provides a spectrum sensing method for a three-dimensional broadband spectrum sensing communication system, which can solve the problems of decreased detection performance and insufficient generalization ability in complex channel environments, long-distance low signal-to-noise ratio conditions, and dynamic spectrum changes.

[0006] To achieve the above objectives, according to a first aspect of this application, an adaptive neural network for three-dimensional broadband spectrum sensing is provided, comprising an initial convolutional layer module, a dual prior encoder module, a space-frequency time-decoupled convolutional network module, a feature linear modulation module, a multi-head self-attention mechanism module, and a global pooling layer module; the initial convolutional layer module receives three-dimensional broadband spectrum sensing variables and outputs time-domain and amplitude variation features; the dual prior encoder module receives three-dimensional broadband spectrum sensing variables and outputs a comprehensive prior vector; the space-frequency time-decoupled convolutional network module receives time-domain and amplitude variation features and outputs a decoupled feature tensor; the feature linear modulation module receives the comprehensive prior vector and the decoupled feature tensor and outputs the feature linear modulation module output; the multi-head self-attention mechanism module receives the feature linear modulation module output and outputs attention-enhanced features; the global pooling layer module receives the attention-enhanced features and outputs a broadband spectrum occupancy probability prediction value and a normalized signal-to-noise ratio prediction value, respectively.

[0007] To achieve the above objectives, according to a second aspect of this application, a spectrum sensing method for a three-dimensional broadband spectrum sensing communication system is also provided. The three-dimensional broadband spectrum sensing communication system includes several ground master users and a sensing UAV. The UAV communicates via... An array of unit antennas receives broadband signals from a ground primary user and estimates the occupancy state and signal-to-noise ratio for each subband of the broadband signal; the method uses an adaptive neural network for three-dimensional broadband spectrum sensing as described in the first aspect; the method includes the following steps: S1. The three-dimensional broadband spectrum sensing variables are transmitted to the initial convolutional layer module, and the temporal and amplitude variation characteristics are output. S2. The three-dimensional broadband spectrum sensing variables are transmitted to the dual prior encoder module, which outputs a comprehensive prior vector. S3. Transmit the temporal and amplitude variation features to the space-frequency decoupled convolutional network module and output the decoupled feature tensor. S4. Transmit the combined prior vector and the decoupled feature tensor to the feature linear modulation module, and output the output of the feature linear modulation module. S5. Transmit the output of the feature linear modulation module to the multi-head self-attention mechanism module to output attention-enhanced features; S6. The attention enhancement features are transmitted to the global pooling layer module, which outputs the broadband spectrum occupancy probability prediction value and the normalized signal-to-noise ratio prediction value, respectively.

[0008] Furthermore, the channel model of the three-dimensional broadband spectrum sensing communication system includes line-of-sight propagation, non-line-of-sight propagation, shadowing fading, and multipath fading; the horizontal flight distance between the UAV and the ground-based primary user is set. Drone flight altitude 3D slant distance horizontal distance of flight Flight altitude of drones Angle of elevation The path loss under the line-of-sight condition is then... The path loss under non-line-of-sight conditions is The signals received by the secondary user include line-of-sight and non-line-of-sight signals, with the probability of receiving a line-of-sight signal being... Based on the probability of receiving line-of-sight Path loss under line-of-sight conditions Path loss under non-line-of-sight conditions The overall path loss is obtained. Let the total channel bandwidth be According to the total channel bandwidth Calculate receiver noise power Combined with comprehensive path loss The theoretical signal-to-noise ratio (SNR) of each spectral sub-band is obtained, i.e., the sub-band SNR. ,in, Transmit power for primary users.

[0009] Furthermore, the three-dimensional broadband spectrum sensing variables include broadband signals transmitted by the ground primary user; S1 further includes the following steps: Drones The single-element antenna array receives broadband signals from the ground primary user and combines them into a broadband complex signal sampling sequence. ,in, The number of subbands to be divided for the spectrum bandwidth For the number of antennas, Number of time sampling points; wideband complex signal sampling sequence After unfolding the real and imaginary parts, the initial convolutional layer module is used to extract the low-level features of the broadband signal, and the time-domain and amplitude variation features are obtained by the following formula:

[0010] in, It represents the characteristics of time-domain and amplitude variations.

[0011] Furthermore, the three-dimensional broadband spectrum sensing variables also include those derived from the flight horizontal distance. Drone flight altitude and elevation angle Triples The triplet With wideband complex signal sampling sequence Both inputs are fed into the dual-prior encoder module; S2 also includes the following steps: Triple group Transformed into high-dimensional Fourier features with periodic structure enhancement capabilities through Fourier feature mapping:

[0012] The high-dimensional Fourier features are sequentially input into the first fully connected network to obtain low-dimensional geometric embedding vectors. The first fully connected network includes two fully connected layers. The first layer compresses the high-dimensional Fourier features to 64 dimensions to extract the main components, and the second layer further performs linear transformation to improve the separability of geometric features. Both fully connected layers are used in conjunction with nonlinear activation functions to generate low-dimensional geometric embedding vectors with 64-dimensional expressive power. Based on the broadband complex signal sampling sequence Constructing a channel quality statistical feature set And generate channel quality vectors through a second fully connected network. The second fully connected network has the same structure as the first fully connected network. Low-dimensional geometric embedding vector With channel quality vector The comprehensive prior vector is obtained by concatenating the following formula:

[0013] in, This is a comprehensive prior vector.

[0014] Furthermore, S3 also includes the following steps: The space-frequency decoupled convolutional network module extracts spatial correlation, subcarrier frequency domain coupling, and temporal dynamic changes respectively, and constructs three convolutional branches that do not share parameters, including spatial convolutional branch, frequency domain convolutional branch, and temporal convolutional branch; Spatial convolution branch Used to extract spatial correlation between multiple antenna arrays: Frequency domain convolution branch Extract local frequency domain coupling features on each sub-band: Temporal convolution branch Modeling the dynamic changes in the received signal over time: ; The outputs of the three branch convolutional branches are fused in the channel dimension to generate a decoupled feature tensor. .

[0015] Furthermore, step S4 also includes the following steps: The characteristic linear modulation module integrates prior vectors As input, the scaling factor is generated using the following formula. and bias factor :

[0016] in, Scaling factor The weight matrix, Scaling factor The bias vector, Bias factor The weight matrix, Bias factor The bias vector; scaling factor and bias factor Applying this to the convolution output yields the output of the characteristic linear modulation module. :

[0017] in, For Hadama accumulation.

[0018] Furthermore, S5 also includes the following steps: The multi-head self-attention mechanism module will use the output features of the feature linear modulation module. The query is performed using the following formula. ,key ,value Linear transformation:

[0019] The multi-head self-attention mechanism includes There are 1 attention head, and the output of each attention head is calculated using the following formula:

[0020] in, For key Feature dimensions; Will The outputs of each attention head are concatenated by channel dimension: Then through the weight matrix Transformation to obtain attention-enhanced features .

[0021] Furthermore, S6 also includes the following steps: The global pooling layer module enhances attention features. The vector is transmitted to the global pooling operator to obtain the joint feature representation vector. :

[0022] in, This represents the global average pooling operator applied to both the event and subband dimensions; Joint feature representation vector The signals are fed into the broadband spectrum occupancy probability output branch and the signal-to-noise ratio regression output branch, respectively. The broadband spectrum occupancy probability output branch is as follows:

[0023] in, For the first Predicted broadband spectrum occupancy probability values ​​for each sub-band It is the sigmoid activation function. To occupy the detection branch The weight vector of each sub-band To occupy the detection branch The bias term of each sub-band; the signal-to-noise ratio regression output branch is:

[0024] in, For the first Normalized signal-to-noise ratio prediction for each sub-band It is a non-linear function that takes the maximum value element by element. For the signal-to-noise ratio regression output branch The transpose of the weight vectors of each sub-band. For the signal-to-noise ratio regression output branch The bias term of each sub-band.

[0025] Furthermore, S6 also includes the following steps: The consistency regularization term is constructed using the following formula. :

[0026] in, The number of subbands to be divided for the spectrum bandwidth This is a linear normalization function for the signal-to-noise ratio regression results in the [0,1] interval; Consistency regularization The final joint loss function is obtained through the following formula:

[0027] in, For the binary cross-entropy loss of the spectrum occupancy detection task, This represents the mean squared error loss for the signal-to-noise ratio regression task. These are the weighting coefficients for the signal-to-noise ratio regression loss. The weighting coefficients for the consistency regularization term are... and Used to balance the contributions of each loss term.

[0028] In this application, a dual-prior encoder module fully utilizes the geometric position information and received signal statistical characteristics of the UAV, enabling the network to dynamically adjust the feature extraction method according to the flight position and channel conditions, thereby maintaining stable perception capabilities in non-stationary channels. A space-frequency-time decoupled convolutional network module effectively characterizes the spatial correlation of multiple antennas, sub-band frequency coupling relationships, and temporal dynamic characteristics in broadband signals, improving the precision and discriminative power of feature representation. The introduction of a feature linear modulation module allows for channel-by-channel linear adjustment of the network's internal features based on prior information, making the model more effective at capturing weak signal features in low signal-to-noise ratio scenarios and avoiding over-response to high-energy regions in high signal-to-noise ratio scenarios, thus significantly improving weak signal recognition performance and overall robustness. The multi-head self-attention mechanism module models the global dynamic relationship between broadband subcarriers, automatically focusing on the frequency components that have the greatest impact on spectrum state decisions. This compensates for the insufficient ability of traditional convolutional structures to represent long-distance correlations across subbands, further improving the decision accuracy of broadband spectrum sensing in complex interference environments. A joint output structure integrating spectrum occupancy detection and signal-to-noise ratio estimation is constructed, enabling the model to not only determine whether the spectrum is occupied but also simultaneously output the channel quality index of the corresponding subband, thereby improving the physical interpretability of the detection results. The consistency regularization constraint introduced during training ensures that the occupancy probability and signal-to-noise ratio prediction remain consistent under weak signal conditions, effectively reducing the false alarm rate and improving the sensing reliability in long-distance, weak-signal environments. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 This is a diagram of an adaptive neural network structure for three-dimensional broadband spectrum sensing provided according to an embodiment of this application; Figure 2 This is a flowchart of a spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to an embodiment of this application; Figure 3 This is a three-dimensional geometric relationship diagram between the UAV and the ground main user according to an embodiment of this application; Figure 4 This is a comparison chart of detection probability curves for different methods provided in the embodiments of this application as the horizontal flight distance changes. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] like Figure 1 As shown, this application provides an adaptive neural network for three-dimensional broadband spectrum sensing, including an initial convolutional layer module, a dual prior encoder module, a space-frequency decoupled convolutional network module, a feature linear modulation module, a multi-head self-attention mechanism module, and a global pooling layer module. The initial convolutional layer module receives three-dimensional broadband spectrum sensing variables and outputs temporal and amplitude variation features. The dual prior encoder module receives three-dimensional broadband spectrum sensing variables and outputs a comprehensive prior vector. The space-frequency decoupled convolutional network module receives temporal and amplitude variation features and outputs a decoupled feature tensor. The feature linear modulation module receives the comprehensive prior vector and the decoupled feature tensor and outputs the feature linear modulation module output. The multi-head self-attention mechanism module receives the feature linear modulation module output and outputs attention-enhanced features. The global pooling layer module receives the attention-enhanced features and outputs a broadband spectrum occupancy probability prediction value and a normalized signal-to-noise ratio prediction value, respectively.

[0033] like Figure 2 As shown, this application also provides a spectrum sensing method for a three-dimensional broadband spectrum sensing communication system, the three-dimensional broadband spectrum sensing communication system including several ground master users and a sensing UAV, the UAV through... The unit antenna array receives broadband signals from the ground primary user and estimates the occupancy state and signal-to-noise ratio for each subband of the broadband signal; the method uses the aforementioned adaptive neural network for three-dimensional broadband spectrum sensing; the method includes the following steps: S1. The three-dimensional broadband spectrum sensing variables are transmitted to the initial convolutional layer module, and the temporal and amplitude variation characteristics are output. S2. The three-dimensional broadband spectrum sensing variables are transmitted to the dual prior encoder module, which outputs a comprehensive prior vector. S3. Transmit the temporal and amplitude variation features to the space-frequency decoupled convolutional network module and output the decoupled feature tensor. S4. Transmit the combined prior vector and the decoupled feature tensor to the feature linear modulation module, and output the output of the feature linear modulation module. S5. Transmit the output of the feature linear modulation module to the multi-head self-attention mechanism module to output attention-enhanced features; S6. The attention enhancement features are transmitted to the global pooling layer module, which outputs the broadband spectrum occupancy probability prediction value and the normalized signal-to-noise ratio prediction value, respectively.

[0034] like Figure 3 As shown, this application further considers a typical urban air-to-ground communication scenario to characterize the spatial geometric relationship between the UAV and the ground primary users and its impact on path loss and signal-to-noise ratio. The urban air-to-ground communication scenario includes several ground primary users and a sensing UAV, which communicates via... The unit antenna array receives the broadband signal from the ground primary user and estimates the occupancy state and signal-to-noise ratio for each subband of the broadband signal; the channel model for the urban air-to-ground communication scenario includes line-of-sight (LoS), non-line-of-sight (NLoS) propagation, shadow fading, and multipath fading; assuming the horizontal flight distance between the UAV and the ground primary user is... The drone's flight altitude is Then the three-dimensional slant distance Calculated using the following formula:

[0035] horizontal distance of flight Flight altitude of drones Angle of elevation Calculated using the following formula:

[0036] As the drone's trajectory changes, the horizontal flight distance... 3D slant distance and elevation angle All of these factors exhibit dynamic changes, resulting in significant three-dimensional spatial non-stationarity of channel characteristics. This application simultaneously considers the superposition effects of free space loss, additional fading under line-of-sight and non-line-of-sight conditions, and shadow fading; the path loss under the line-of-sight condition... This can be expressed by the following formula:

[0037] in, For carrier frequency, At the speed of light, The shadowing fading term of the log-normal distribution. The environmentally relevant line-of-sight additional loss constant; Path loss under non-line-of-sight conditions This can be expressed by the following formula:

[0038] in, This is the non-line-of-sight additional loss constant. This is a small-scale fading term.

[0039] For air-to-ground communication in urban environments, the signals received by secondary users include line-of-sight and non-line-of-sight signals, where the probability of receiving line-of-sight signals is... This can be expressed by the following formula:

[0040] in, and These are all environmental parameters, determined by factors such as building density and height, and the received line-of-sight probability is... The probability that the received signal is at line of sight; based on the probability that the received signal is at line of sight. The overall path loss is obtained. :

[0041] Overall path loss The distance between the drone and the ground node increases monotonically with increasing horizontal distance.

[0042] Regarding the derivation of the signal-to-noise ratio, let the total channel bandwidth be... Receiver noise power It is calculated by the following formula:

[0043] in, For thermal noise power spectral density, For receiver noise figure; combined with integrated path loss The theoretical signal-to-noise ratio (SNR) of each spectral sub-band is obtained, i.e., the sub-band SNR. :

[0044] in, Main user transmit power; due to three-dimensional slant range Horizontal distance of flight The overall path loss increases with the increase in size. The signal-to-noise ratio of the subband increases with distance, eventually leading to an increase in the signal-to-noise ratio of the subband. Horizontal distance of flight The signal-to-noise ratio of the subband increases but then decreases. This application will... After normalization, it is used as one of the prior features of the network to guide subsequent deep models, enabling them to have stronger geometric perception and weak signal adaptability.

[0045] Furthermore, the three-dimensional broadband spectrum sensing variables include broadband signals transmitted by the ground primary user.

[0046] Furthermore, S1 also includes the following steps: Drones The single-element antenna array receives broadband signals from the ground primary user and combines them into a broadband complex signal sampling sequence. ,in, The number of subbands to be divided for the spectrum bandwidth For the number of antennas, Number of time sampling points; wideband complex signal sampling sequence After unfolding the real and imaginary parts, the initial convolutional layer module is used to extract the low-level features of the broadband signal, and the time-domain and amplitude variation features are obtained by the following formula:

[0047] in, It represents the characteristics of time-domain and amplitude variations.

[0048] Furthermore, the three-dimensional broadband spectrum sensing variables also include those derived from the flight horizontal distance. Drone flight altitude and elevation angle Triples The triplet With wideband complex signal sampling sequence Both are input to the dual prior encoder module.

[0049] Furthermore, S2 also includes the following steps: Triple group Transformed into high-dimensional Fourier features with periodic structure enhancement capabilities through Fourier feature mapping:

[0050] The high-dimensional Fourier features are sequentially input into the first fully connected network to obtain low-dimensional geometric embedding vectors. The first fully connected network includes two fully connected layers. The first layer compresses the high-dimensional Fourier features to 64 dimensions to extract the main components, and the second layer further performs linear transformation to improve the separability of geometric features. Both fully connected layers are used in conjunction with nonlinear activation functions to generate low-dimensional geometric embedding vectors with 64-dimensional expressive power. Based on the broadband complex signal sampling sequence Constructing a channel quality statistical feature set And generate channel quality vectors through a second fully connected network. The second fully connected network has the same structure as the first fully connected network. Low-dimensional geometric embedding vector With channel quality vector The comprehensive prior vector is obtained by concatenating the following formula:

[0051] in, This is a comprehensive prior vector.

[0052] Furthermore, S3 also includes the following steps: The space-frequency decoupled convolutional network module extracts spatial correlation, subcarrier frequency domain coupling, and temporal dynamic changes respectively, and constructs three convolutional branches that do not share parameters, including spatial convolutional branch, frequency domain convolutional branch, and temporal convolutional branch; Spatial convolution branch Used to extract spatial correlation between multiple antenna arrays: Frequency domain convolution branch Extract local frequency domain coupling features on each sub-band: Temporal convolution branch Modeling the dynamic changes in the received signal over time: ; The outputs of the three branch convolutional branches are fused in the channel dimension to generate a decoupled feature tensor. .

[0053] Furthermore, step S4 also includes the following steps: The characteristic linear modulation module integrates prior vectors As input, the scaling factor is generated using the following formula. and bias factor :

[0054] in, Scaling factor The weight matrix, Scaling factor The bias vector, Bias factor The weight matrix, Bias factor The bias vector; scaling factor and bias factor Applying this to the convolution output yields the output of the characteristic linear modulation module. :

[0055] in, For Hadama accumulation.

[0056] Furthermore, S5 also includes the following steps: The multi-head self-attention mechanism module will use the output features of the feature linear modulation module. The query is performed using the following formula. ,key ,value Linear transformation:

[0057] The multi-head self-attention mechanism includes There are 1 attention head, and the output of each attention head is calculated using the following formula:

[0058] in, For key Feature dimensions; Will The outputs of each attention head are concatenated by channel dimension: Then through the weight matrix Transformation to obtain attention-enhanced features .

[0059] Furthermore, S6 also includes the following steps: The global pooling layer module enhances attention features. The vector is transmitted to the global pooling operator to obtain the joint feature representation vector. :

[0060] in, This represents the global average pooling operator applied to both the event and subband dimensions; Joint feature representation vector The signals are fed into the broadband spectrum occupancy probability output branch and the signal-to-noise ratio regression output branch, respectively. The broadband spectrum occupancy probability output branch is as follows:

[0061] in, For the first Predicted broadband spectrum occupancy probability values ​​for each sub-band It is the sigmoid activation function. To occupy the detection branch The weight vector of each sub-band To occupy the detection branch The bias term of each sub-band; the signal-to-noise ratio regression output branch is:

[0062] in, For the first Normalized signal-to-noise ratio prediction for each sub-band It is a non-linear function that takes the maximum value element by element. For the signal-to-noise ratio regression output branch The transpose of the weight vectors of each sub-band. For the signal-to-noise ratio regression output branch The bias term of each sub-band.

[0063] Furthermore, S6 also includes the following steps: The consistency regularization term is constructed using the following formula. :

[0064] in, The number of subbands to be divided for the spectrum bandwidth This is a linear normalization function for the signal-to-noise ratio regression results in the [0,1] interval; Consistency regularization The final joint loss function is obtained through the following formula:

[0065] in, For the binary cross-entropy loss of the spectrum occupancy detection task, This represents the mean squared error loss for the signal-to-noise ratio regression task. These are the weighting coefficients for the signal-to-noise ratio regression loss. The weighting coefficients for the consistency regularization term are... and Used to balance the contributions of each loss term.

[0066] In one specific embodiment, regarding system deployment and data acquisition, this application considers a drone performing a low-altitude inspection mission in an urban area. The drone maintains a flight altitude of approximately 100m and a horizontal flight distance varying between 100m and 1000m. The drone is equipped with a three-antenna linear array receiver, which samples a 2.4GHz broadband signal at a fixed sampling rate. The sampled complex baseband signal is organized into a format with a size of [missing information]. The time-series input contains multiple spectral sub-bands, and a unified processing flow is used to complete sub-band occupancy detection and signal-to-noise ratio estimation. The UAV positioning module outputs the flight horizontal distance, UAV flight altitude, and elevation angle calculated from them in real time. These parameters, together with channel statistics, are used to construct prior features to enhance the model's perception capabilities in dynamic environments.

[0067] Regarding the generation process of prior features, the UAV in this application continuously generates geometric prior features in real time based on the position parameters provided by the flight control system during flight. Specifically, the flight horizontal distance, UAV flight altitude and attitude parameters provided by the positioning module are first mapped into three-dimensional geometric quantities, and then input into a lightweight fully connected network after Fourier feature mapping to obtain low-dimensional geometric embeddings. Meanwhile, the envelope statistics of the received signal (including average power, variance, amplitude distribution, etc.) are calculated by the front-end DSP module and then input into the channel prior encoder to generate a channel quality vector. The two types of priors are concatenated in the channel dimension to obtain a comprehensive prior vector, which provides modulation guidance in the subsequent network inference stage. This part can be executed by the flight control SoC or a separate MCU in engineering deployment, running in parallel with the main network without affecting the actual inference speed. The DSP module is a digital signal processing unit deployed in the UAV receiver hardware, used to preprocess and extract statistical features of the broadband complex signal. After data sampling, it performs operations such as power calculation, variance analysis, amplitude distribution statistics, and noise filtering on the input signal to generate a channel statistical feature set. The channel quality vector is generated by inputting the channel prior encoder.

[0068] For the network inference process, the complex baseband signals acquired by the three antennas are first unfolded according to the time sequence and input into the initial convolutional layer to obtain the basic feature representation. Subsequently, the features are fed into a three-branch convolutional structure in the spatial, frequency, and time domains to extract array directional features, frequency coupling features, and signal temporal dynamics, respectively. The fused features are then modulated channel-by-channel by a feature linear modulation module based on real-time prior vectors to adapt to different channel conditions. The modulated features are then input into a multi-head self-attention unit to extract the global coupling relationship between subcarriers. Finally, after global pooling, two output heads are input: the first outputs the occupancy probability of the spectral subband, and the second outputs the predicted signal-to-noise ratio (SNR) value of the corresponding subband. This embodiment uses a threshold decision method to obtain the final subband occupancy status. This process enables continuous, real-time broadband spectrum sensing during UAV flight.

[0069] For model training and testing methods, this application generates multiple training samples under different distances and positions based on simulated air-to-ground channels. The training set covers a horizontal flight distance range from 100m to 1000m, while the test set additionally includes previously unseen distance points (such as 450m, 650m, and 850m) to verify the model's generalization ability. During training, a joint loss function is used to achieve collaborative optimization of the occupancy detection task and the signal-to-noise ratio regression task. To enhance robustness under weak signal conditions, a consistency regularization term constraint is introduced between the occupancy probability and the signal-to-noise ratio prediction, making the decision results more stable. In the testing phase, the model is deployed on a GPU server for offline inference evaluation, and compared with commonly used methods (such as energy detection, ParallelCNN, DeepSense, and Transformer) to verify the performance advantages of this application under weak signal and long-distance conditions.

[0070] Table 1 lists the parameter size, number of floating-point operations (FLOPs), and single inference time for different methods, illustrating the computational complexity and resource consumption of each model in actual deployment. As can be seen from Table 1, the traditional Energy Detection (ED) method, due to its independence from deep models, has the shortest inference time and lowest computational cost, but it cannot utilize multi-dimensional features, limiting its accuracy in complex air-to-ground channel environments. Lightweight networks such as ParallelCNN and DeepSense have advantages in parameter size and inference speed, but their structures are relatively fixed and lack the ability to adaptively model priors and channel states in three-dimensional space. While Transformer networks have strong feature representation capabilities, their parameter count and computational cost are significantly higher than other models, leading to a significant increase in inference time and making them unsuitable for real-time deployment on resource-constrained UAV platforms. The method proposed in this application introduces prior modulation and attention mechanisms into the network structure while maintaining a moderate number of parameters and an acceptable number of floating-point operations, keeping computational complexity and real-time performance within a deployable range, providing a better balance between performance and complexity for engineering applications.

[0071] Table 1. Computational complexity and resource consumption of each model

[0072] As can be seen from Table 1, the model size and computational cost of this application are within a reasonable range, which is better than the computational cost of high-complexity models and can realize real-time inference on embedded UAV platforms, thus making it feasible for practical deployment.

[0073] Figure 4 The diagram illustrates the detection probability curves of different methods as a function of horizontal flight distance. The horizontal axis represents the horizontal flight distance between the UAV and the ground user, and the vertical axis represents the average detection probability. As the horizontal flight distance increases, the path loss of the air-to-ground link increases, and the signal-to-noise ratio decreases. The detection performance of all methods declines, reflecting the non-stationary nature of real-world 3D air-to-ground channels. Traditional methods (such as ED, DeepSense, and ParallelCNN) show significant performance degradation at long distances and in low signal-to-noise ratio regions. While Transformer possesses stronger feature representation capabilities, its adaptability to spatial geometric changes under weak signal conditions is limited. This application maintains higher detection probabilities across multiple distance ranges, exhibiting a significant advantage, particularly in the low signal-to-noise ratio region of 800-1000m. This demonstrates that by introducing geometric priors, channel quality priors, feature linear modulation, and attention mechanisms, this application effectively enhances the network's ability to capture weak signal features and improves detection stability and robustness under long-distance links. Figure 4This further verifies that the present application can maintain high detection accuracy and environmental adaptability in a three-dimensional air-ground environment, and its performance under long-distance and weak signal conditions is significantly better than the comparative method.

[0074] This application has the following beneficial effects: This application introduces three-dimensional geometric position priors and channel quality priors into a broadband spectrum sensing framework, constructing an adaptive neural network structure capable of adapting to complex air-to-ground channel variations. Compared to existing technologies, it exhibits significant performance and application advantages. Traditional deep learning spectrum sensing methods are mostly based on planar or static channel assumptions, lacking the ability to model changes in the three-dimensional spatial structure of the UAV flight environment. Especially under long-distance or weak signal conditions, spectrum occupancy features are easily masked by noise, leading to a significant decrease in detection performance. In contrast, this application fully utilizes the UAV's geometric position information and received signal statistical characteristics through a dual-prior encoder module, enabling the network to dynamically adjust the feature extraction method according to the flight position and channel conditions, thereby maintaining stable sensing capabilities in non-stationary channels.

[0075] In terms of feature extraction, this application employs a space-frequency time-decoupled convolutional network module, which effectively characterizes the spatial correlation of multiple antennas, sub-band frequency coupling relationship, and temporal dynamic characteristics in broadband signals, thereby improving the precision and discriminative ability of feature representation. Furthermore, the introduction of a feature linear modulation module enables the linear adjustment of the network features channel by channel based on prior information, making it easier for the model to capture weak signal features in low signal-to-noise ratio scenarios and avoid over-response to strong energy regions in high signal-to-noise ratio scenarios, thus significantly improving the weak signal recognition performance and overall robustness.

[0076] This application models the global dynamic relationship between broadband subcarriers through a multi-head self-attention mechanism module, which can automatically focus on the frequency components that have the greatest impact on the spectrum state decision, compensate for the problem that traditional convolutional structures are insufficient in representing long-distance correlations across subbands, and further improve the decision accuracy of broadband spectrum sensing in complex interference environments.

[0077] This application constructs a joint output structure integrating spectrum occupancy detection and signal-to-noise ratio estimation, enabling the model to not only determine whether the spectrum is occupied, but also to simultaneously output the channel quality index of the corresponding sub-band, thereby improving the physical interpretability of the detection results. The consistency regularization constraint introduced during the training process ensures that the occupancy probability and signal-to-noise ratio prediction remain consistent under weak signal conditions, effectively reducing the false alarm rate and improving the perception reliability in long-distance and weak signal environments.

[0078] Overall, this application achieves high-precision and robust detection capability of broadband spectrum in dynamic three-dimensional air-ground channel environment by introducing a prior information-driven adaptive learning mechanism, a space-frequency time-decoupled convolutional network module, a feature linear modulation module, and a multi-head self-attention mechanism module. It is significantly better than existing spectrum sensing methods based on fixed channel models or single feature extraction methods, and has good engineering application value and promotion prospects.

[0079] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various steps of the aforementioned three-dimensional broadband spectrum sensing method embodiment based on prior information adaptive learning, and can achieve the same beneficial effects as the aforementioned three-dimensional broadband spectrum sensing method embodiment based on prior information adaptive learning. To avoid repetition, these will not be described again here.

[0080] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0081] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0082] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An adaptive neural network for three-dimensional broadband spectrum sensing, characterized in that, It includes an initial convolutional layer module, a dual prior encoder module, a space-frequency time-decoupled convolutional network module, a feature linear modulation module, a multi-head self-attention mechanism module, and a global pooling layer module. The initial convolutional layer module receives three-dimensional broadband spectrum sensing variables and outputs temporal and amplitude variation features. The dual prior encoder module receives three-dimensional broadband spectrum sensing variables and outputs a comprehensive prior vector. The space-frequency time-decoupled convolutional network module receives temporal and amplitude variation features and outputs a decoupled feature tensor. The feature linear modulation module receives the integrated prior vector and the decoupled feature tensor, and outputs the feature linear modulation module output; the multi-head self-attention mechanism module receives the feature linear modulation module output and outputs attention-enhanced features. The global pooling layer module receives attention-enhanced features and outputs predicted values ​​for broadband spectrum occupancy probability and normalized signal-to-noise ratio, respectively.

2. A spectrum sensing method for a three-dimensional broadband spectrum sensing communication system, the three-dimensional broadband spectrum sensing communication system comprising several ground master users and a sensing UAV, the UAV communicating via... The unit antenna array receives broadband signals from ground primary users and estimates the occupancy status and signal-to-noise ratio for each subband of the broadband signal; characterized in that, The method uses the adaptive neural network for three-dimensional broadband spectrum sensing as described in claim 1; the method includes the following steps: S1. The three-dimensional broadband spectrum sensing variables are transmitted to the initial convolutional layer module, and the temporal and amplitude variation characteristics are output. S2. The three-dimensional broadband spectrum sensing variables are transmitted to the dual prior encoder module, which outputs a comprehensive prior vector. S3. Transmit the temporal and amplitude variation features to the space-frequency decoupled convolutional network module and output the decoupled feature tensor. S4. Transmit the combined prior vector and the decoupled feature tensor to the feature linear modulation module, and output the output of the feature linear modulation module. S5. Transmit the output of the feature linear modulation module to the multi-head self-attention mechanism module to output attention-enhanced features; S6. The attention enhancement features are transmitted to the global pooling layer module, which outputs the broadband spectrum occupancy probability prediction value and the normalized signal-to-noise ratio prediction value, respectively.

3. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to claim 2, characterized in that, The channel model of the three-dimensional broadband spectrum sensing communication system includes line-of-sight propagation, non-line-of-sight propagation, shadowing fading, and multipath fading; the horizontal flight distance between the UAV and the ground main user is set. Drone flight altitude 3D slant distance horizontal distance of flight Flight altitude of drones Angle of elevation The path loss under the line-of-sight condition is then... The path loss under non-line-of-sight conditions is The signals received by the secondary user include line-of-sight and non-line-of-sight signals, with the probability of receiving a line-of-sight signal being... Based on the probability of receiving line-of-sight Path loss under line-of-sight conditions Path loss under non-line-of-sight conditions The overall path loss is obtained. Let the total channel bandwidth be According to the total channel bandwidth Calculate receiver noise power Combined with comprehensive path loss The theoretical signal-to-noise ratio (SNR) of each spectral sub-band is obtained, i.e., the sub-band SNR. ,in, Transmit power for primary users.

4. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to claim 3, characterized in that, The three-dimensional broadband spectrum sensing variables include broadband signals transmitted by ground primary users; S1 further includes the following steps: Drones The single-element antenna array receives broadband signals from the ground primary user and combines them into a broadband complex signal sampling sequence. ,in, The number of subbands to be divided for the spectrum bandwidth For the number of antennas, Number of time sampling points; wideband complex signal sampling sequence After unfolding the real and imaginary parts, the initial convolutional layer module is used to extract the low-level features of the broadband signal, and the time-domain and amplitude variation features are obtained by the following formula: in, It represents the characteristics of time-domain and amplitude variations.

5. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to claim 4, characterized in that, The three-dimensional broadband spectrum sensing variables also include those derived from flight horizontal distance. Drone flight altitude and elevation angle Triples The triplet With wideband complex signal sampling sequence Both inputs are fed into the dual-prior encoder module; S2 also includes the following steps: Triple group Transformed into high-dimensional Fourier features with periodic structure enhancement capabilities through Fourier feature mapping: The high-dimensional Fourier features are sequentially input into the first fully connected network to obtain low-dimensional geometric embedding vectors. The first fully connected network includes two fully connected layers. The first layer compresses the high-dimensional Fourier features to 64 dimensions to extract the main components, and the second layer further performs linear transformation to improve the separability of geometric features. Both fully connected layers are used in conjunction with nonlinear activation functions to generate low-dimensional geometric embedding vectors with 64-dimensional expressive power. Based on the broadband complex signal sampling sequence Constructing a channel quality statistical feature set And generate channel quality vectors through a second fully connected network. The second fully connected network has the same structure as the first fully connected network. Low-dimensional geometric embedding vector With channel quality vector The comprehensive prior vector is obtained by concatenating the following formula: in, This is a comprehensive prior vector.

6. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to claim 5, characterized in that, S3 also includes the following steps: The space-frequency decoupled convolutional network module extracts spatial correlation, subcarrier frequency domain coupling, and temporal dynamic changes respectively, and constructs three convolutional branches that do not share parameters, including spatial convolutional branch, frequency domain convolutional branch, and temporal convolutional branch; Spatial convolution branch Used to extract spatial correlation between multiple antenna arrays: Frequency domain convolution branch Extract local frequency domain coupling features on each sub-band: Temporal convolution branch Modeling the dynamic changes in the received signal over time: ; The outputs of the three branch convolutional branches are fused in the channel dimension to generate a decoupled feature tensor. .

7. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to claim 6, characterized in that, S4 also includes the following steps: The characteristic linear modulation module integrates prior vectors As input, the scaling factor is generated using the following formula. and bias factor : in, Scaling factor The weight matrix, Scaling factor The bias vector, Bias factor The weight matrix, Bias factor The bias vector; scaling factor and bias factor Applying this to the convolution output yields the output of the characteristic linear modulation module. : in, For Hadama accumulation.

8. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to claim 7, characterized in that, S5 also includes the following steps: The multi-head self-attention mechanism module will use the output features of the feature linear modulation module. The query is performed using the following formula. ,key ,value Linear transformation: The multi-head self-attention mechanism includes There are 1 attention head, and the output of each attention head is calculated using the following formula: in, For key Feature dimensions; Will The outputs of each attention head are concatenated by channel dimension: Then through the weight matrix Transformation to obtain attention-enhanced features .

9. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to claim 8, characterized in that, S6 also includes the following steps: The global pooling layer module enhances attention features. The vector is transmitted to the global pooling operator to obtain the joint feature representation vector. : in, This represents the global average pooling operator applied to both the event and subband dimensions; Joint feature representation vector The signals are fed into the broadband spectrum occupancy probability output branch and the signal-to-noise ratio regression output branch, respectively. The broadband spectrum occupancy probability output branch is as follows: in, For the first Predicted broadband spectrum occupancy probability values ​​for each sub-band It is the sigmoid activation function. To occupy the detection branch The weight vector of each sub-band To occupy the detection branch The bias term of each sub-band; the signal-to-noise ratio regression output branch is: in, For the first Normalized signal-to-noise ratio prediction for each sub-band It is a non-linear function that takes the maximum value element by element. For the signal-to-noise ratio regression output branch The transpose of the weight vectors of each sub-band. For the signal-to-noise ratio regression output branch The bias term of each sub-band.

10. The spectrum sensing method for a three-dimensional broadband spectrum sensing communication system according to any one of claims 2-9, characterized in that, S6 also includes the following steps: The consistency regularization term is constructed using the following formula. : in, The number of subbands to be divided for the spectrum bandwidth This is a linear normalization function for the signal-to-noise ratio regression results in the [0,1] interval; Consistency regularization The final joint loss function is obtained through the following formula: in, For the binary cross-entropy loss of the spectrum occupancy detection task, This represents the mean squared error loss for the signal-to-noise ratio regression task. These are the weighting coefficients for the signal-to-noise ratio regression loss. The weighting coefficients for the consistency regularization term are... and Used to balance the contributions of each loss term.

Images