Unmanned aerial vehicle detection method based on radio frequency spectrum identification and deep learning

Abstract

The invention discloses an unmanned aerial vehicle detection method based on radio frequency spectrum identification and deep learning, and relates to the technical field of unmanned aerial vehicle monitoring, and the method comprises the steps: deploying a radio receiving station to collect radio signals, converting the radio signals into a time-frequency graph, extracting multi-scale signal features, carrying out the global average pooling of the multi-scale signal features, and carrying out the global average pooling of the multi-scale signal features; calculating a scale adjustment coefficient to generate a weighted feature, extracting a time sequence signal feature to generate a fusion feature, constructing a full connection layer to perform multi-task processing on the fusion feature, and outputting an unmanned aerial vehicle detection result; a state space is defined, an action space is defined according to environmental parameters, nonlinear harmonic response characteristics in a current state are measured to construct a nonlinear reward function, a comprehensive reward function is formed in combination with action interference effects, and a final interference strategy is output through reinforcement learning. According to the invention, the scientificity and robustness of the identification and interference strategy are significantly improved, the accuracy of the interference strategy is improved, and the interference effect of the unmanned aerial vehicle is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) monitoring technology, and in particular to a UAV detection method based on radio spectrum identification and deep learning. Background Technology

[0002] In recent years, with the increasing popularity of drone applications, drone detection technology based on radio spectrum has gradually become an important means of security monitoring and spectrum management. Early related technologies were mostly based on traditional signal detection and feature extraction methods, using energy detection, correlation detection, and time-domain feature analysis to perform simple identification of radio signals in the field. These methods are simple and intuitive, but they are easily affected by complex environmental interference, multipath effects, and signal concealment. The accuracy and robustness of identification are difficult to meet the needs of practical applications. As the utilization rate of spectrum space continues to increase, simple energy threshold detection is no longer sufficient to distinguish the spectrum characteristics occupied by drones, prompting research to gradually shift to frequency domain and time-frequency domain analysis, combined with pattern recognition technology to improve recognition performance. However, existing technologies still face many challenges in practical applications. Traditional machine learning models based on shallow features are sensitive to signal-to-noise ratio and cannot fully utilize the deep information in complex signals, which restricts the accuracy and robustness of drone detection. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a drone detection method based on radio spectrum identification and deep learning, which solves the problem that existing traditional machine learning models based on shallow features are sensitive to signal-to-noise ratio and have difficulty fully utilizing deep information in complex signals, thus limiting the accuracy and robustness of drone detection.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for detecting unmanned aerial vehicles (UAVs) based on radio spectrum identification and deep learning, comprising, A radio receiving station is deployed to collect radio signals, which are then converted into time-frequency maps. A multi-scale convolutional neural network is constructed to extract multi-scale signal features. Global average pooling is performed on the multi-scale signal features, and scale adjustment coefficients are calculated to generate weighted features. Simultaneously, time-series signal features are extracted through a multi-layer LSTM network and concatenated with the weighted features to generate fused features. A fully connected layer is constructed to perform multi-task processing on the fused features and output the UAV detection results. Collect basic environmental parameters, construct a multivariate spatiotemporal prediction model to predict future spectrum occupancy, generate a spectrum situation matrix, and further construct a safe interference space; The state space is defined based on the UAV detection results, and the action space is defined based on the environmental parameters. The nonlinear harmonic response characteristics of the current state are measured to construct a nonlinear reward function. The comprehensive reward function is formed by combining the action interference effect. The comprehensive reward function is optimized through reinforcement learning and the final interference strategy is output.

[0006] As a preferred embodiment of the UAV detection method based on radio spectrum identification and deep learning described in this invention, the step of converting radio signals into time-frequency graphs, constructing multi-scale convolutional neural networks to extract multi-scale signal features, performing global average pooling on the multi-scale signal features, and calculating scale adjustment coefficients to generate weighted features refers to using short-time Fourier transform to convert radio signals into time-frequency graphs. A multi-scale convolutional neural network is constructed using three parallel branches, each using a convolutional kernel of a different size to extract multi-scale signal features; Multi-scale features are obtained by splicing signal features of different scales along the channel direction. ; Global average pooling of multi-scale features is computed using an adaptive attention mechanism; The coefficient s is adjusted based on the global average pooling generation scale; Weighted features are generated by enhancing multi-scale features using a scaling adjustment coefficient. .

[0007] As a preferred embodiment of the UAV detection method based on radio spectrum identification and deep learning described in this invention, the step of extracting time-series signal features and weighted features through a multi-layer LSTM network and concatenating them to generate fused features, and constructing a fully connected layer to perform multi-task processing on the fused features and output the UAV detection results refers to using a two-layer LSTM network to construct a multi-layer LSTM model, wherein the number of hidden units in each LSTM layer is set to 128, and the output of the first layer is passed to the second layer. The weighted features are concatenated with the time-series signal features extracted by the LSTM model to form fused features. ; The fused features are processed in multiple tasks through a fully connected layer, including drone category determination, signal type identification, carrier frequency point regression, bandwidth regression, modulation method classification, signal strength and signal-to-noise ratio regression. The results of multi-task processing are merged and output as the UAV detection results.

[0008] As a preferred embodiment of the UAV detection method based on radio spectrum identification and deep learning described in this invention, the following steps are taken: collecting basic environmental parameters, constructing a multivariate spatiotemporal prediction model to predict future spectrum occupancy, generating a spectrum situation matrix, and further constructing a security interference spatial index. This is achieved by collecting the current time t, receiver coordinates G, and target frequency from the radio receiving station. and legal frequency bands Define the spectrum state input vector ; Extract historical signal data from the database to obtain historical spectrum occupancy data. ; A multivariate spatiotemporal prediction model was constructed and trained using a multidimensional graph neural network. The model input included historical spectrum occupancy data. and spectrum state input vector The output shows the future spectrum occupancy status. ; Frequency points are extracted from future spectrum occupancy trends. In the future Occupancy probability ; A spectrum situation matrix is ​​constructed using the predicted occupancy probabilities, and a set of frequency points where interference is permissible is defined. ; Constructing a secure interference space based on frequency point sets .

[0009] As a preferred embodiment of the UAV detection method based on radio spectrum identification and deep learning described in this invention, the step of defining the state space based on UAV detection results and defining the action space according to environmental parameters refers to extracting the UAV category and signal-to-noise ratio based on the UAV detection results, and extracting the available interference frequency points, maximum interference power, and target frequency points from the safe interference space to form the state space. ; Based on the target frequency Define interference frequency points ; Set the interference bandwidth according to the bandwidth range of the radio receiving station. ; Adjusting the jamming transmission power based on the maximum jamming power ; Interference frequency points Interference bandwidth and interference transmission power Composition of action space .

[0010] As a preferred embodiment of the UAV detection method based on radio spectrum identification and deep learning described in this invention, the nonlinear harmonic response characteristics measured in the current state are used to construct a nonlinear reward function, which is then combined with the action interference effect to form a comprehensive reward function. The comprehensive reward function is optimized through reinforcement learning, and the final interference strategy is output from the time-frequency diagram. In the process of identifying peak values, the frequency corresponding to the maximum peak value is taken as the key frequency point. Extract the amplitude of the corresponding key frequency points from the time-frequency graph. Based on the amplitude of key frequency points, the energy of radio signals at the second and third harmonic frequencies is extracted; Reward bias terms are constructed based on the energy of radio signals at second and third harmonic frequencies. ; Simultaneously construct multiple reward functions based on interference effects and legal frequency bands. ; A comprehensive reward function is constructed by integrating multiple reward functions with reward bias terms. ; The comprehensive reward function is used as the optimization objective function. A policy network and a value network are used to optimize the objective function, and the optimized actions are output through the policy network. The minimum mean square error is calculated based on the comprehensive reward function. The action that minimizes the mean square error is output as the optimal action, and the final interference strategy is formed based on the optimal action.

[0011] As a preferred embodiment of the UAV detection method based on radio spectrum identification and deep learning described in this invention, the final interference strategy is then synchronously stored in the database along with the collected UAV signals.

[0012] As a preferred embodiment of the UAV detection method based on radio spectrum identification and deep learning described in this invention, wherein: the deployment of a radio receiving station to collect radio signals refers to the real-time collection of UAV radio signals through the radio receiving station, with the sampling frequency set to [missing information]. ; The raw sampled radio signal from the radio receiving station is over-gain calibrated; A digital FIR bandpass filter is used to perform spectral filtering on the gain-calibrated radio signal. The filtered radio signal is frequency-shifted to baseband. The radio signal after frequency conversion is normalized.

[0013] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the UAV detection method based on radio spectrum identification and deep learning as described in the first aspect of the present invention.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the UAV detection method based on radio spectrum identification and deep learning as described in the first aspect of the present invention.

[0015] The beneficial effects of this invention are as follows: This invention extracts rich frequency domain features through multi-scale deep convolutional neural networks and combines them with multi-layer LSTM networks for time-series modeling, realizing synchronous analysis of spectrum space and time dynamics, significantly improving the scientificity and robustness of identification and interference strategies. Furthermore, by combining nonlinear harmonic features with the interference effect of the strategy, and using deep learning networks to optimize the interference strategy, the accuracy of the interference strategy is improved, effectively enhancing the interference effect of UAVs. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the UAV detection method based on radio spectrum identification and deep learning in Example 1. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a drone detection method based on radio spectrum identification and deep learning, including the following steps: S1. Deploy a radio receiving station to collect radio signals, convert the radio signals into time-frequency maps, and construct a multi-scale convolutional neural network to extract multi-scale signal features. Perform global average pooling on the multi-scale signal features, calculate the scale adjustment coefficient to generate weighted features, and simultaneously extract time-series signal features through a multi-layer LSTM network and concatenate them with the weighted features to generate fused features. Construct a fully connected layer to perform multi-task processing on the fused features and output the UAV detection results. Specifically, deploying a radio receiving station to collect radio signals refers to collecting the radio signals of the drone in real time through a radio receiving station, setting the sampling frequency to [missing information]. : in This is the highest sampling frequency; The raw sampled radio signal from the radio receiving station was over-gain calibrated: in The original sampled radio signal, The gain parameters for the radio receiving station are obtained in advance through calibration experiments. This is the radio signal after gain calibration; The gain-calibrated radio signal is subjected to spectral filtering using a digital FIR bandpass filter: in and These are the Fast Fourier Transform and the Inverse Fourier Transform, respectively. For bandpass filters, This is the filtered radio signal; Shift the filtered radio signal to baseband: Where j is the imaginary unit. The local oscillator frequency is set to the target center frequency of the monitoring signal. is the radio signal after frequency conversion, and n is the signal sequence; The radio signal after frequency conversion is normalized.

[0022] The highest sampling frequency ensures comprehensive acquisition of spectrum information, avoiding omissions and playing a decisive role in the accurate capture of UAV communication signals in complex environments. Reasonable gain calibration compensates for variations in the front-end transmission link, improving signal amplitude consistency and making subsequent frequency domain feature extraction more stable, thus enhancing recognition robustness. Utilizing an FIR bandpass filter to filter noise in non-target frequency bands effectively extracts valid signals from the target frequency band, providing clear and centralized frequency domain features for the detection algorithm. Frequency shifting to baseband technology solves the carrier frequency deviation problem, making the spectrum features more stable and suitable for deep learning model input, and also contributing to fast and accurate UAV signal recognition. Signal normalization significantly reduces the inconsistency of different signal amplitudes, improving the efficiency of subsequent feature extraction, deep learning model training, and interference strategy formulation, while also greatly enhancing recognition accuracy and environmental adaptability.

[0023] Furthermore, the radio signal is converted into a time-frequency graph, and a multi-scale convolutional neural network is constructed to extract multi-scale signal features. Global average pooling is then performed on the multi-scale signal features to calculate scale adjustment coefficients and generate weighted features. Short-time Fourier transform is then used to convert the radio signal into a time-frequency graph. Where n is the radio signal number. It is a Hanning window, with a length of It is applied at time t, where j is the imaginary unit. The sampling frequency of the radio receiving station is f, which is a frequency variable and covers the main frequency band of the signal. A multi-scale convolutional neural network is constructed using three parallel branches, each using convolutional kernels of different sizes to extract multi-scale signal features: in For the signal features at the i-th scale, and Let be the convolution kernel and bias for the i-th scale. For convolution operators, in this invention, the convolution scales respectively include , as well as ; Multi-scale features are obtained by splicing signal features of different scales along the channel direction. : in , as well as Signal features at different scales; Global average pooling of multi-scale features is computed using an adaptive attention mechanism: Where H and W are the feature space dimensions, and C is the number of channels. Let z be the feature value of the c-th channel at position i, k, and z be the result of global pooling. Adjust the coefficient s based on the global average pooling generation scale: in Here is the weight matrix of the fully connected layer, with a size of C×C. For bias terms, Use the Sigmoid activation function; Weighted features are generated by enhancing multi-scale features using a scaling adjustment coefficient. : in For each channel, multiply.

[0024] By using time-frequency maps, the instantaneous frequency changes and energy distribution of a signal can be captured simultaneously, which helps in identifying complex or hidden drone signals, especially in scenarios with environmental noise interference or diverse signal modulation. Employing a Hanning window with a specific length can reduce spectral leakage, improve spectral resolution, and ensure higher accuracy of extracted signal features. Convolutional kernels of different scales can extract features sensitive to different bandwidths and local structures, especially in complex and variable spectral environments, enabling the capture of key signal features from multiple angles and scales. This step is a crucial means to overcome the limitations of traditional single-scale feature extraction and improve the robustness of model recognition. Designing different scales separately ensures a good response to both subtle changes and macroscopic features in the signal, greatly improving recognition accuracy. By stitching together multi-scale and multi-angle features, a rich feature representation is formed, which helps subsequent attention mechanisms or fully connected layers extract important information more effectively. This step overcomes the drawback of isolated information at different scales, maximizing feature resources. This approach enhances the model's ability to distinguish complex signals. Global average pooling yields representative values ​​for each channel, suppressing local noise and highlighting the global characteristics of the signal. An adaptive attention mechanism is introduced, dynamically weighting features at different scales to emphasize key scales or important features, effectively improving the model's insensitivity to changes in different signals. Adjustment coefficients can specifically control the weights of features at different scales in subsequent enhancement, achieving dynamic enhancement or suppression of signal features. This effectively addresses changing signal environments, improving the model's adaptability and robustness. The generated coefficients are multiplied by the corresponding features channel by channel to selectively enhance features at specific scales. The model automatically adjusts the emphasis on features at different scales based on actual environmental changes, enhancing the ability to capture significant signals and suppressing the impact of interference or noise, thereby improving detection capabilities in severely interfered environments.

[0025] Furthermore, by extracting temporal signal features and concatenating them with weighted features through a multi-layer LSTM network to generate fused features, a fully connected layer is constructed to perform multi-task processing on the fused features and output the UAV detection results. This refers to using a two-layer LSTM network to construct a multi-layer LSTM model, where the number of hidden units in each LSTM layer is set to 128, and the output of the first layer is passed to the second layer. in and For the first and second layers of LSTM, The length of the signal sequence. The output of the first layer LSTM network at time t. The output of the second-layer LSTM network at time t is used as the feature of the time-series signal. The weighted features are concatenated with the time-series signal features extracted by the LSTM model to form fused features. : The fused features are processed in multiple tasks through a fully connected layer, including drone category determination, signal type identification, carrier frequency point regression, bandwidth regression, modulation method classification, signal strength and signal-to-noise ratio regression. In multi-task processing, tasks such as drone category determination, signal type identification, and modulation method classification need to output results through activation functions, for example: in For the i-th task result, including the drone category Signal type and modulation method , and These are the weights and biases of the i-th task, respectively; The remaining tasks output weights and biases: in The result of the i-th task includes carrier frequency point regression. Bandwidth regression Signal strength RSSI and signal-to-noise ratio regression SNR; The results of multi-task processing are merged and output as the UAV detection results.

[0026] By extracting features layer by layer, the model's understanding of long-term dependencies is deepened, solving the problem of difficult signal feature extraction in dynamic scenes, thereby improving the robustness of UAV recognition. In applications, especially in interference environments or under multipath conditions, these deep features are more stable and discriminative. By adjusting the importance of different frequency bands or features through a weighting mechanism, it helps to more accurately distinguish UAV signals in weak signals or complex interference backgrounds, strengthening the emphasis on key features for specific tasks and greatly improving signal recognition performance. Especially in the presence of multi-source interference, by combining recognition categories and physical parameter regression, a multi-dimensional compatible recognition system is formed, effectively solving the problem of "incomplete recognition of a single task" and avoiding the phenomenon of feature silos. The model can share basic features during the learning process, strengthening the connection and synergy between different tasks, thereby improving the accuracy of multi-target tasks while reducing model complexity.

[0027] S2. Collect basic environmental parameters, construct a multivariate spatiotemporal prediction model to predict future spectrum occupancy, generate a spectrum situation matrix, and further construct a safe interference space. Specifically, basic environmental parameters are collected, a multivariate spatiotemporal prediction model is constructed to predict future spectrum occupancy, a spectrum situation matrix is ​​generated, and a security interference spatial index is further constructed. The model is based on data collected from the radio receiving station, including current time t, receiving station coordinates G, and target frequency. and legal frequency bands Define the spectrum state input vector : Extract historical signal data from the database to obtain historical spectrum occupancy data. : in Let i be the power level at the i-th frequency point at time t-1. For time intervals; A multivariate spatiotemporal prediction model was constructed and trained using a multidimensional graph neural network. The model input included historical spectrum occupancy data. and spectrum state input vector The output shows the future spectrum occupancy status. : Frequency points are extracted from future spectrum occupancy trends. In the future Occupancy probability : in For scale parameters, For bias terms, For frequency point In the future The predicted occupancy feature vector is extracted from the future spectrum occupancy trend; A spectrum situation matrix is ​​constructed using the predicted occupancy probabilities, and a set of frequency points where interference is permissible is defined. : in This is the probability threshold; Constructing a secure interference space based on frequency point sets : in The maximum interference power at the frequency point. The maximum interference power as defined by regulations.

[0028] The innovation of using a multidimensional graph neural network model, which combines historical spectrum occupancy data with current environmental parameters, lies in its ability to fully consider the interaction between spatial relationships and temporal evolution within the spectrum. This model can capture the transmission characteristics and temporal dependencies of signals at different spatial locations, solving the problem of insufficient prediction accuracy of single temporal or spatial models in complex environments. Using this model, the likelihood of future frequency occupancy can be anticipated, enabling early spectrum resource coordination and interference space optimization, effectively avoiding blind interference or unexpected interference. The spectrum situation matrix intuitively reflects the future spectrum resource utilization trend, exhibiting a high degree of foresight. By filtering high-occupancy-probability frequencies through probability thresholds, a "set of permissible interference frequencies" is constructed. This process greatly reduces blind interference behavior and avoids accidentally damaging legitimate frequency bands, thereby maximizing interference efficiency while protecting legitimate use. Using future spectrum occupancy prediction results, the interference space is dynamically defined, ensuring both the success rate of interference targets and compliance with legal and regulatory restrictions. The maximum interference power limit design ensures the rationality and safety of the interference intensity. The multivariate spatiotemporal prediction model achieves efficient prediction of future spectrum space, not only enhancing the accuracy and legality of interference but also significantly improving the intelligence level of the entire detection and interference system.

[0029] S3. Define the state space based on the UAV detection results and the action space based on the environmental parameters. Measure the nonlinear harmonic response characteristics of the current state to construct a nonlinear reward function. Combine the action interference effect to form a comprehensive reward function. Optimize the comprehensive reward function through reinforcement learning and output the final interference strategy. Specifically, defining the state space based on UAV detection results and the action space based on environmental parameters involves extracting the UAV category and signal-to-noise ratio from the UAV detection results, and extracting available interference frequencies, maximum interference power, and target frequencies from the safe interference space to form the state space. : Based on the target frequency Define interference frequency points : in The network output adjustment parameters are in the range of [range missing]. , The interference frequency range is determined by changes in the frequency domain environment and the required interference accuracy. Set the interference bandwidth according to the bandwidth range of the radio receiving station. : in and These represent the maximum and minimum bandwidth output values ​​of the radio receiving station. This is the bandwidth output adjustment parameter, with a range of [range missing]. ; Adjusting the jamming transmission power based on the maximum jamming power : in This is the power output adjustment parameter, with a range of [range missing]. ; Interference frequency points Interference bandwidth and interference transmission power Composition of action space .

[0030] By adjusting network output parameters and dynamically adjusting the interference frequency range, precise "point-to-point" interference can be located, greatly improving the interference success rate. Utilizing environmental parameter constraints and combined with observed spectrum changes, the interference bandwidth and transmission power are automatically adjusted. This mechanism ensures that the interference has sufficient frequency domain coverage without exceeding the limits, avoiding side effects caused by illegal interference or spectrum chaos. Integrating frequency, bandwidth, and power into the action space gives the system a higher degree of control freedom. This design breaks through the limitations of linear adjustment, forming a multi-modal, multi-objective optimization space. It can scientifically and quickly adjust the interference strategy according to environmental changes. The action space provides rich adjustment parameters. Through intelligent algorithms such as reinforcement learning, the system can autonomously learn environmental changes and optimize adjustment strategies, thereby maintaining the interference effect in complex signal environments.

[0031] Furthermore, a nonlinear reward function is constructed by measuring the nonlinear harmonic response characteristics under the current state. This is combined with the action interference effect to form a comprehensive reward function. The comprehensive reward function is then optimized using reinforcement learning, and the final interference strategy is output from the time-frequency diagram. In the process of identifying peak values, the frequency corresponding to the maximum peak value is taken as the key frequency point. Extract the amplitude of the corresponding key frequency points from the time-frequency graph. Based on the amplitude of key frequency points, the energy of radio signals at the second and third harmonic frequencies is extracted: in and These are the key frequencies of radio signals at corresponding frequency points. and The amplitude reflects the energy of the radio signal at the second and third harmonic frequencies. and The energy of radio signals at second and third harmonic frequencies; Reward bias terms are constructed based on the energy of radio signals at second and third harmonic frequencies. : in As a reward adjustment coefficient, and To adjust the weighting of the target harmonic deviation in the reward, and The target harmonic amplitude at the interference frequency is obtained through calculation at the target frequency. Simultaneously construct multiple reward functions based on interference effects and legal frequency bands. : in The jamming success criterion is 1 if the drone deviates from its route or communication is interrupted at the jamming frequency point; otherwise, it is 0. This is a signal penalty term. If the interfering frequency is within a legal frequency band, the penalty value is 1; otherwise, it is 0. To interfere with invalid indicators, , as well as These are the weighting coefficients; The Related to the success rate of interference, it is calculated using the following formula: A comprehensive reward function is constructed by integrating multiple reward functions with reward bias terms. : in This is the adjustment coefficient; The comprehensive reward function is used as the optimization objective function, which is then optimized using a policy network and a value network. The input to the policy network is the state. The output is an action. The value network input is the state. The output is a state value estimate, which is then used by the policy network to output optimized actions. The minimum mean square error is calculated based on the comprehensive reward function. The action that minimizes the mean square error is output as the optimal action, and the final interference strategy is formed based on the optimal action.

[0032] In complex spectra, the energy peaks of a signal are representative. Peak detection allows for rapid and accurate determination of target frequencies, effectively avoiding noise interference and random fluctuations from interference sources in the interference spectrum. This improves the accuracy of target frequency identification and lays the foundation for subsequent energy analysis. Utilizing second- and third-order harmonic energy can yield the nonlinear "fingerprint" of a signal, providing deep information that traditional linear features struggle to access. This significantly enhances the underlying physical basis of identification. Compared to using only amplitude or linear spectra, this feature significantly improves the identification of unknown signals and anti-interference capabilities, effectively solving the identification challenges in environments with spectral distortion and multi-source interference. The deviation reward mechanism defines the deviation value based on the target harmonic characteristics. The guiding model focuses more on the nonlinear characteristics of energy changes. This design can significantly optimize the training process, make up for the problem of insufficient expression of nonlinear system characteristics by traditional models, accelerate the convergence speed, and ensure the generalization ability of the model. Combining interference success, signal penalty and legal frequency band indicators, a multi-objective reward mechanism is designed to effectively solve the problem of blind interference under a single objective. This allows the system to maximize the interference effect while complying with regulations and avoiding legal risks. By using deep strategies and value networks, the target is gradually approached, effectively responding to the dynamic changes of the environmental spectrum. The training objective of minimum mean square error ensures the continuity and optimality of the strategy output, greatly improving the stability and targeting of the interference effect.

[0033] Furthermore, after the final jamming strategy is formed, it is stored in the database in sync with the collected drone signals.

[0034] This embodiment also provides a computer device applicable to the drone detection method based on radio spectrum identification and deep learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the drone detection method based on radio spectrum identification and deep learning proposed in the above embodiment.

[0035] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0036] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the UAV detection method based on radio spectrum identification and deep learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0037] In summary, this invention extracts rich frequency domain features through multi-scale deep convolutional neural networks and combines them with multi-layer LSTM networks for time-series modeling, achieving synchronous analysis of the spectrum space and time dynamics. This significantly improves the scientific rigor and robustness of the identification and interference strategies. Furthermore, by combining nonlinear harmonic features with the interference effect of the strategies and using deep learning networks to optimize the interference strategies, the accuracy of the interference strategies is improved, effectively enhancing the interference effect of UAVs.

[0038] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting unmanned aerial vehicles (UAVs) based on radio spectrum identification and deep learning, characterized in that: include, A radio receiving station is deployed to collect radio signals, which are then converted into time-frequency maps. A multi-scale convolutional neural network is constructed to extract multi-scale signal features. Global average pooling is performed on the multi-scale signal features, and scale adjustment coefficients are calculated to generate weighted features. Simultaneously, time-series signal features are extracted through a multi-layer LSTM network and concatenated with the weighted features to generate fused features. A fully connected layer is constructed to perform multi-task processing on the fused features and output the UAV detection results. Collect basic environmental parameters, construct a multivariate spatiotemporal prediction model to predict future spectrum occupancy, generate a spectrum situation matrix, and further construct a safe interference space; The state space is defined based on the UAV detection results, and the action space is defined based on the environmental parameters. The nonlinear harmonic response characteristics of the current state are measured to construct a nonlinear reward function. The comprehensive reward function is formed by combining the action interference effect. The comprehensive reward function is optimized through reinforcement learning and the final interference strategy is output.

2. The UAV detection method based on radio spectrum identification and deep learning as described in claim 1, characterized in that: The process of converting radio signals into time-frequency graphs, constructing multi-scale convolutional neural networks to extract multi-scale signal features, performing global average pooling on the multi-scale signal features, and calculating scale adjustment coefficients to generate weighted features refers to using short-time Fourier transform to convert radio signals into time-frequency graphs. A multi-scale convolutional neural network is constructed using three parallel branches, each using a convolutional kernel of a different size to extract multi-scale signal features; Multi-scale features are obtained by splicing signal features of different scales along the channel direction. ; Global average pooling of multi-scale features is computed using an adaptive attention mechanism; The coefficient s is adjusted based on the global average pooling generation scale; Weighted features are generated by enhancing multi-scale features using a scaling adjustment coefficient. .

3. The UAV detection method based on radio spectrum identification and deep learning as described in claim 2, characterized in that: The step of extracting time-series signal features and weighted features through a multi-layer LSTM network to generate fused features, and constructing a fully connected layer to perform multi-task processing on the fused features to output UAV detection results refers to using a two-layer LSTM network to construct a multi-layer LSTM model, wherein the number of hidden units in each LSTM layer is set to 128, and the output of the first layer is passed to the second layer. The weighted features are concatenated with the time-series signal features extracted by the LSTM model to form fused features. ; The fused features are processed in multiple tasks through a fully connected layer, including drone category determination, signal type identification, carrier frequency point regression, bandwidth regression, modulation method classification, signal strength and signal-to-noise ratio regression. The results of multi-task processing are merged and output as the UAV detection results.

4. The UAV detection method based on radio spectrum identification and deep learning as described in claim 3, characterized in that: The collected basic environmental parameters are used to construct a multivariate spatiotemporal prediction model to predict future spectrum occupancy trends, generating a spectrum trend matrix. Further, a security interference spatial index is constructed by collecting data from the radio receiving station, including current time t, receiving station coordinates G, and target frequency. and legal frequency bands Define the spectrum state input vector ; Extract historical signal data from the database to obtain historical spectrum occupancy data. ; A multivariate spatiotemporal prediction model was constructed and trained using a multidimensional graph neural network. The model input included historical spectrum occupancy data. and spectrum state input vector The output shows the future spectrum occupancy status. ; Frequency points are extracted from future spectrum occupancy trends. In the future Occupancy probability ; A spectrum situation matrix is ​​constructed using the predicted occupancy probabilities, and a set of frequency points where interference is permissible is defined. ; Constructing a secure interference space based on frequency point sets .

5. The UAV detection method based on radio spectrum identification and deep learning as described in claim 4, characterized in that: The process of defining the state space based on UAV detection results and the action space based on environmental parameters refers to extracting the UAV category and signal-to-noise ratio from the UAV detection results, and extracting available interference frequencies, maximum interference power, and target frequencies from the safe interference space to form the state space. ; Based on the target frequency Define interference frequency points ; Set the interference bandwidth according to the bandwidth range of the radio receiving station. ; Adjusting the jamming transmission power based on the maximum jamming power ; Interference frequency points Interference bandwidth and interference transmission power Composition of action space .

6. The UAV detection method based on radio spectrum identification and deep learning as described in claim 5, characterized in that: The nonlinear harmonic response characteristics measured in the current state are used to construct a nonlinear reward function. This function is then combined with the action interference effect to form a comprehensive reward function. Reinforcement learning is used to optimize the comprehensive reward function and output the final interference strategy based on the time-frequency diagram. In the process of identifying peak values, the frequency corresponding to the maximum peak value is taken as the key frequency point. Extract the amplitude of the corresponding key frequency points from the time-frequency graph. Based on the amplitude of key frequency points, the energy of radio signals at the second and third harmonic frequencies is extracted; Reward bias terms are constructed based on the energy of radio signals at second and third harmonic frequencies. ; Simultaneously construct multiple reward functions based on interference effects and legal frequency bands. ; A comprehensive reward function is constructed by integrating multiple reward functions with reward bias terms. ; The comprehensive reward function is used as the optimization objective function. A policy network and a value network are used to optimize the objective function, and the optimized actions are output through the policy network. The minimum mean square error is calculated based on the comprehensive reward function. The action that minimizes the mean square error is output as the optimal action, and the final interference strategy is formed based on the optimal action.

7. The UAV detection method based on radio spectrum identification and deep learning as described in claim 6, characterized in that: After the final jamming strategy is formed, it is stored in the database in sync with the collected UAV signals.

8. The UAV detection method based on radio spectrum identification and deep learning as described in claim 1, characterized in that: The deployment of the radio receiving station to collect radio signals refers to the real-time collection of radio signals from the UAV via the radio receiving station, with the sampling frequency set to [frequency value missing]. ; The raw sampled radio signal from the radio receiving station is over-gain calibrated; A digital FIR bandpass filter is used to perform spectral filtering on the gain-calibrated radio signal. The filtered radio signal is frequency-shifted to baseband. The radio signal after frequency conversion is normalized.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the UAV detection method based on radio spectrum identification and deep learning as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the UAV detection method based on radio spectrum identification and deep learning as described in any one of claims 1 to 8.

Images