An unmanned aerial vehicle intelligent detection method based on opportunity signal

Abstract

The application provides an unmanned aerial vehicle intelligent detection method based on opportunity signals, belongs to the technical field of target detection in the radio passive sensing domain, and solves the problems such as active detection exposure of existing unmanned aerial vehicle monitoring technologies in complex outdoor environments. The method comprises the following steps: collecting opportunity signal samples in the presence and absence of unmanned aerial vehicles respectively, collecting all opportunity signal samples and corresponding unmanned aerial vehicle presence state labels, and constructing an unmanned aerial vehicle perception dataset; taking the complex baseband sequence of each opportunity signal sample in the unmanned aerial vehicle perception dataset as the model input, taking the unmanned aerial vehicle presence state label of the corresponding opportunity signal sample as the output label, training the unmanned aerial vehicle intelligent detection model, and obtaining the trained unmanned aerial vehicle intelligent detection model; receiving real-time opportunity signals, inputting the complex baseband sequence of the real-time opportunity signals into the trained unmanned aerial vehicle intelligent detection model, and predicting the output to obtain the detection result of whether the unmanned aerial vehicle exists.

Description

Technical Field

[0001] This invention relates to the field of target detection technology in the passive sensing domain of radio, and in particular to an intelligent detection method for unmanned aerial vehicles (UAVs) based on opportunistic signals. Background Technology

[0002] With the rapid popularization of the low-altitude economy and drone applications, drones have played an important role in various fields such as aerial surveying and mapping, logistics inspection, and emergency rescue. However, the problems of illegal intrusion, unauthorized flights, and "black flights" they bring also pose significant security risks to airport airspace, perimeter security of important facilities, and security for major events. Therefore, developing a drone monitoring technology that can be deployed long-term, is cost-effective, and remains robust even in adversarial environments has become a key research direction in the field of low-altitude security. Existing drone detection technologies mainly include radar detection, photoelectric / infrared imaging, acoustic detection, and radio detection based on drone communication links. Among them, radar and photoelectric detection methods are prone to false alarms and missed alarms under complex terrain backgrounds, obstructions, and multipath conditions, and achieving stable detection of low RCS miniaturized drones remains a challenge; acoustic methods have limited detection range and are easily affected by environmental noise; communication link-based detection methods rely on the active launch behavior and protocol of drones, which are easily evaded in silent or adversarial conditions, and their adaptability across drone models and scenarios is still insufficient.

[0003] In contrast, signals of opportunity offer advantages such as wide coverage, continuous and stable transmission, and abundant spectrum resources. Passive UAV detection using signals of opportunity can achieve covert monitoring without actively radiating signals, offering advantages like low power consumption, low detectability, and ease of deployment, making it particularly suitable for perimeter security and long-term monitoring scenarios. However, in open outdoor environments, the propagation of signals of opportunity is significantly affected by terrain, vehicles, pedestrians, and building reflections, resulting in strong multipath propagation, non-stationary noise, co-channel interference, and dynamic background clutter. UAV-scattered echoes are typically weak and easily drowned out by background noise. This makes it difficult for traditional methods based on fixed thresholds, energy detection, or a limited number of manually statistical features to stably characterize the subtle time-frequency disturbances caused by UAVs, generally exhibiting sensitivity to environmental changes and insufficient generalization ability. Furthermore, while deep learning has shown outstanding performance in automatic feature extraction and complex pattern modeling in recent years, providing a new technical path for signal-based UAV detection, most existing deep learning solutions directly model the original baseband sequence, easily learning scene-related incidental features, leading to performance degradation across regions and time.

[0004] Therefore, how to solve the problems of existing drone monitoring technology in complex outdoor environments, such as the vulnerability of active detection, poor environmental adaptability due to reliance on human features, and insufficient cross-scenario generalization ability, has become a key issue that urgently needs to be addressed. Summary of the Invention

[0005] This invention relates to an intelligent detection method for unmanned aerial vehicles (UAVs) based on opportunistic signals, aiming to solve the problems of existing UAV monitoring technologies in complex outdoor environments, such as the vulnerability of active detection, poor environmental adaptability due to reliance on human features, and insufficient generalization ability across scenarios.

[0006] This invention provides an intelligent detection method for unmanned aerial vehicles (UAVs) based on opportunistic signals, characterized in that the method includes: Opportunity signal samples were collected in scenarios with and without drones, and all opportunity signal samples and their corresponding drone presence status labels were aggregated to construct a drone perception dataset. A drone intelligent detection model is constructed by sequentially connecting a learnable filter bank and a drone detection network. The complex baseband sequence of each opportunity signal sample in the drone perception dataset is used as the model input, and the drone presence state label of the corresponding opportunity signal sample is used as the output label. The drone intelligent detection model is trained to obtain the trained drone intelligent detection model. The system receives real-time opportunity signals, inputs the complex baseband sequence of the real-time opportunity signals into the trained UAV intelligent detection model, and predicts the output to obtain the detection result of whether the UAV exists.

[0007] Based on the above solution, the present invention also makes the following improvements: Furthermore, during the training process of the UAV intelligent detection model, the joint training loss function of the UAV intelligent detection model... Represented as:

[0008] in, This represents the binary cross-entropy loss function of the drone detection network. , Let represent the frequency regularization function and the diversity regularization function of the learnable filter bank, respectively. , These represent the weight coefficients of the frequency regularization function and the diversity regularization function, respectively.

[0009] Furthermore, the learnable filter bank is implemented by a one-dimensional convolutional layer containing multiple parallel convolutional kernels, each corresponding to an FIR subband filter; each convolutional kernel parameter serves as the tap coefficient of the corresponding FIR subband filter, and the convolutional kernel parameters can be updated during training.

[0010] Furthermore, the frequency regularization function of the filter bank can be learned. Represented as:

[0011] in, , They represent the first The preset passband frequency set and stopband frequency set of each FIR subband filter The preset positive stability constant; Indicates the first Tap coefficients of each FIR subband filter The frequency domain response obtained by performing a discrete Fourier transform. This represents a discrete frequency index variable.

[0012] further,

[0013] in, This represents the length of the FIR subband filter.

[0014] Furthermore, diversity regularization function Represented as:

[0015] in, Indicates the first The, the The correlation matrix between the normalized convolution kernels of the FIR subband filters This indicates the total number of FIR subband filters.

[0016] further,

[0017] in, , They represent the first The, the The normalized convolution kernel parameters of each filter.

[0018] Furthermore, regarding the first Tap coefficients of each FIR subband filter Normalization is performed to obtain the first... Normalized convolution kernels for each FIR subband filter :

[0019] in, It is a stable term.

[0020] Furthermore, the detection result of whether the drone exists is obtained by performing the following operations to predict the output: The complex baseband sequence of real-time opportunity signals is input into the trained UAV intelligent detection model, and the multi-subband signal is predicted and output by the trained learnable filter bank. The multi-channel subband signal is input into the drone detection network, and the drone detection network predicts and outputs the detection result of whether the drone exists.

[0021] Furthermore, the opportunity signal includes one or more of the following: FM broadcast signal, AM broadcast signal, digital audio broadcast signal, digital television broadcast signal, cellular communication downlink signal, wireless local area network signal, and satellite navigation signal.

[0022] Compared with the prior art, the present invention can achieve at least the following beneficial effects: The intelligent detection method for unmanned aerial vehicles (UAVs) based on opportunistic signals provided by this invention has the following beneficial effects: 1. This invention employs a learnable filter bank implemented by one-dimensional convolution, and uses the convolution kernel as the tap coefficient of the FIR subband filter for end-to-end adaptive updating, thereby realizing multi-subband decomposition and discrimination information enhancement of opportunistic signals. This overcomes the problem of poor adaptability of fixed filter banks to changes in spectrum occupancy and interference distribution, and improves the model's cross-scenario generalization ability.

[0023] 2. In the training phase, this invention combines short-time Fourier transform to construct frequency domain constraint regularization terms. Through passband enhancement and stopband suppression, it guides each filter to learn a frequency response with clear frequency band selectivity, thereby making the model more physically interpretable and stable.

[0024] 3. During the training phase, this invention suppresses filter redundancy and bandwidth collapse by using the diversity constraint regularization term of the filter, causing different filters to converge to different or low-overlapping frequency responses, thereby forming complementary sub-band feature coverage and effectively improving detection performance under complex channel conditions.

[0025] 4. This invention utilizes opportunistic signal passive sensing to achieve UAV detection without active radiation, and has engineering advantages such as low detectability, low power consumption, and easy long-term deployment; at the same time, it uses deep networks to perform end-to-end learning of IQ sequences, reducing the need for manual feature design and complex parameter tuning.

[0026] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from the description and drawings, which are particularly pointed out. Attached Figure Description

[0027] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 A flowchart of an intelligent detection method for unmanned aerial vehicles based on opportunistic signals provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] A specific embodiment of the present invention provides an intelligent detection method for unmanned aerial vehicles (UAVs) based on opportunistic signals, the flowchart of which is shown below. Figure 1 The specific details are as follows.

[0031] Step S1: Collect opportunity signal samples in scenarios with and without drones, summarize all opportunity signal samples and corresponding drone presence status labels, and construct a drone perception dataset. Step S2: Construct an intelligent drone detection model consisting of a learnable filter bank and a drone detection network connected sequentially; use the complex baseband sequence of each opportunity signal sample in the drone perception dataset as the model input, and use the drone presence state label of the corresponding opportunity signal sample as the output label to train the intelligent drone detection model and obtain the trained intelligent drone detection model. Step S3: Receive the real-time opportunity signal, input the complex baseband sequence of the real-time opportunity signal into the trained UAV intelligent detection model, and predict the output to obtain the detection result of whether the UAV exists.

[0032] In step S1, the opportunistic signal is a wireless signal that is radiated by an existing radio transmitter within the monitoring area and can be received without the need for a dedicated transmitter. It requires stable coverage and a reasonable time occupancy rate, and must be able to characterize propagation disturbances such as scattering and multipath propagation caused by the drone in the received signal. The opportunistic signal includes one or more combinations of FM radio signals, AM radio signals, digital audio broadcasting (DAB) signals, digital television broadcasting (DVB-T / DTMB) signals, cellular communication downlink signals, wireless local area network (WiFi) signals, satellite navigation (GNSS) signals, and combinations thereof. This embodiment selects FM signals as the opportunistic signal: a receiving antenna is set up indoors at a fixed height of approximately 150cm to receive FM radio signals within a preset frequency band; in one implementation, a broadcast signal with a center frequency of 97.5MHz within the FM broadcast frequency band (87.5MHz–108MHz) is selected for acquisition, with an acquisition bandwidth of 20MHz and a sampling rate of 25MHz. Data collection was conducted at fixed times each day, with each collection session lasting 30 seconds, for a total of 3 days, in order to obtain sample data covering different time periods.

[0033] In the specific implementation process, opportunistic signal samples were collected in the outdoor monitoring area under both scenarios with and without drones. In the scenario with drones, the drones were located at multiple preset positions (e.g., approximately 1m, 2m, 3m, and 4m from the antenna), and the environmental setup and receiving equipment parameters were kept consistent during the data collection process.

[0034] The collected opportunity signal samples are labeled with the drone's presence status (present / absent) as drone presence status labels. By summarizing all opportunity signal samples and their corresponding drone presence status labels from multiple rounds of collection, a drone perception dataset can be constructed.

[0035] In step S2, the real and imaginary parts of the opportunity signal samples are extracted to obtain corresponding in-phase and orthogonal representations. Opportunity signal samples In-phase components and orthogonal components Composed of complex base band sequences Represented as:

[0036] in, , These represent opportunity signal samples. The real part and the imaginary part.

[0037] Preferably, in this embodiment, the learnable filter bank is implemented by a one-dimensional convolutional layer containing multiple parallel convolutional kernels. Each convolutional kernel corresponds to an FIR subband filter, and the parameters of each convolutional kernel serve as the tap coefficients of the corresponding FIR subband filter. The convolutional kernel parameters can be updated during training, and the learnable filter bank is used to output multiple subband signals. In specific implementation, the constructed learnable filter bank... ,in, Indicates the first The tap coefficients of the FIR subband filter are also the first FIR subband filter. Each convolutional kernel parameter, , This is the length of the FIR subband filter; This indicates the total number of FIR subband filters.

[0038] Opportunity signal samples Complex baseband sequence Input a learnable filter bank, the output of the learnable filter bank is the first... The path carries a signal Represented as:

[0039] In this embodiment, the drone detection network can be implemented using a convolutional neural network. Specifically, the convolutional neural network adopts a ResNet18 residual neural network structure, which is composed of multiple cascaded residual modules. Through multi-layer convolution operations and residual connections, it extracts features of the opportunity signal layer by layer, and finally connects to a classification layer to output the judgment result or corresponding probability of the presence of a drone.

[0040] In the model training process of this embodiment, the joint training loss function of the UAV intelligent detection model Represented as:

[0041] in, This represents the binary cross-entropy loss function of the drone detection network. , Let represent the frequency regularization function and the diversity regularization function of the learnable filter bank, respectively. , These represent the weight coefficients of the frequency regularization function and the diversity regularization function, respectively. In practice, these two weight coefficients can be optimized based on actual conditions. In this embodiment, the binary classification cross-entropy loss... Based on the output of the UAV detection network, a learnable filter bank for optimizing the detection network and jointly updating the front end through backpropagation is calculated, along with a frequency regularization function. and diversity regularization function Both are applied to the convolution kernel parameters of the learnable filter bank, and are used to constrain the frequency band selectivity of the filter and the correlation between different filters, respectively.

[0042] By minimizing the joint training loss, the learnable filter bank and the UAV detection network model are iteratively updated, resulting in a trained learnable filter bank and UAV detection network.

[0043] Binary cross-entropy loss function Represented as:

[0044] in, The batch size for a single training batch. Take 1 to . Indicates the first The real drones present a state label for each opportunity signal sample. The value of is only 0 or 1, where 1 indicates that there is a drone target in the opportunity signal sample (positive sample), and 0 indicates that there is no drone target in the opportunity signal sample (negative sample). Indicates the first The predicted probability value output for each opportunity signal sample represents the model's confidence level in determining the presence of a drone in that signal.

[0045] In the specific implementation process, the frequency domain regularization function is determined in the following way. During the training of the above model, frequency domain constraints are introduced into the learnable filter bank. The frequency domain response is obtained by performing a short-time Fourier transform on the learnable filter, and a frequency domain regularization function is constructed based on passband enhancement and stopband suppression to guide each filter to form a filter response with frequency band selectivity.

[0046] Specifically, for the first Tap coefficients of each FIR subband filter Performing a discrete Fourier transform yields the frequency domain response. :

[0047] in, This represents a discrete frequency index variable.

[0048] Based on a preset passband frequency point set and stopband frequency set Construct frequency domain regularization functions for bandpass enhancement and stopband suppression. :

[0049] in, , They represent the first The preset passband frequency set and stopband frequency set of each FIR subband filter This is a preset positive stability constant. In each training iteration, the frequency domain response of each filter is calculated based on the current convolution kernel parameters, and the frequency domain regularization function is further calculated. The weighted values ​​are added to the joint training loss, and the convolution kernel parameters are updated through backpropagation and gradient descent, so that the stopband response gradually decreases and the passband response is relatively enhanced, thereby guiding each filter to form a frequency response with frequency band selectivity.

[0050] Furthermore, in the specific implementation process, the diversity regularization function is determined in the following way. During the end-to-end joint training phase of the learnable filter bank and the UAV detection network, a diversity constraint is introduced to the learnable filter bank during the training phase. A diversity regularization term is constructed by calculating the correlation between different convolutional kernels, so that different filters converge to different frequency responses.

[0051] For the Tap coefficients of each FIR subband filter Normalization is performed to obtain the first... Normalized convolution kernels for each FIR subband filter :

[0052] in, It is a stable term.

[0053] Construct the correlation matrix based on the normalized convolution kernels of different FIR subband filters, the th , The correlation matrix between the normalized convolution kernels of the FIR subband filters Represented as:

[0054] in, , They represent the first The, the The normalized convolution kernel parameters of each filter are used to construct a diversity regularization function. :

[0055] It should be noted that in this embodiment, the frequency domain regularization term is used to constrain individual filters to form well-defined passband and stopband responses, while the diversity regularization term is used to constrain different filters to reduce correlation and frequency band overlap. Together, they enable the filter bank to possess both frequency band selectivity and subband complementarity. During joint training, the diversity regularization function, classification loss, and frequency domain regularization term are used to construct a joint training loss. The convolution kernel parameters are updated through backpropagation, thereby minimizing the diversity regularization term.

[0056] As described above, the UAV intelligent detection model provided in this embodiment does not employ a subband filter bank with fixed parameters. Instead, it uses one-dimensional convolution kernel parameters as FIR filter tap coefficients and performs end-to-end joint optimization with the subsequent UAV detection network. Simultaneously, it introduces frequency domain constraints and filter diversity constraints, enabling different filters to adaptively generate subband responses with frequency selectivity and complementarity. This enhances the model's adaptability to opportunistic signal spectrum variations and complex propagation environments. For opportunistic signal UAV detection tasks, this method introduces inter-filter correlation constraints into a learnable filter bank and performs joint optimization with frequency domain regularization terms. This ensures that different filters not only possess frequency selectivity but also generate low-correlation, complementary subband responses, reducing filter frequency overlap and response collapse issues.

[0057] In step S3, the complex baseband sequence of the real-time opportunity signal is input into the trained UAV intelligent detection model, and the trained learnable filter bank predicts and outputs multiple sub-band signals. These multiple sub-band signals are then input into the UAV detection network, which predicts and outputs the detection result indicating whether the UAV exists. A more specific implementation process is as follows: For newly received opportunity signals, UAV detection is performed. First, the in-phase and orthogonal representations of the opportunity signal are extracted in a manner consistent with the training samples, and these representations are input into the trained learnable filter bank to obtain multiple sub-band signals. Then, these multiple sub-band signals are input into the trained UAV detection network to obtain the class probabilities of UAV presence and absence. Finally, the detection result indicating whether the UAV exists is determined based on these class probabilities.

[0058] It should be noted that the UAV intelligent detection model trained in the above manner can effectively utilize the spectral characteristics inherent in opportunistic signals to achieve robust detection of low-observable UAV targets without relying on dedicated detection signals. Since the learnable filter bank adaptively optimizes the subband division strategy during training, its output multi-subband signals not only retain the key discriminative information of the original signal but also enhance robustness to different propagation conditions and interference scenarios through frequency band selectivity and complementarity. Furthermore, the end-to-end joint optimization mechanism ensures consistency between front-end filtering and back-end classification tasks, avoiding performance loss caused by target inconsistency in traditional phased designs. Therefore, in practical deployments, this model can be directly applied to real-time RF data streams acquired by the receiver front-end. Only standardized in-phase / quadrature demodulation processing is required to complete the entire inference process from the original signal to the detection decision, demonstrating good engineering practicality and generalization ability.

[0059] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent detection of unmanned aerial vehicles (UAVs) based on opportunistic signals, characterized in that, The method includes: Opportunity signal samples were collected in scenarios with and without drones, and all opportunity signal samples and their corresponding drone presence status labels were aggregated to construct a drone perception dataset. A drone intelligent detection model is constructed by sequentially connecting a learnable filter bank and a drone detection network. The complex baseband sequence of each opportunity signal sample in the drone perception dataset is used as the model input, and the drone presence state label of the corresponding opportunity signal sample is used as the output label. The drone intelligent detection model is trained to obtain the trained drone intelligent detection model. The system receives real-time opportunity signals, inputs the complex baseband sequence of the real-time opportunity signals into the trained UAV intelligent detection model, and predicts the output to obtain the detection result of whether the UAV exists.

2. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to claim 1, characterized in that, During the training of the UAV intelligent detection model, the joint training loss function of the UAV intelligent detection model is... Represented as: in, This represents the binary cross-entropy loss function of the drone detection network. , Let represent the frequency regularization function and the diversity regularization function of the learnable filter bank, respectively. , These represent the weight coefficients of the frequency regularization function and the diversity regularization function, respectively.

3. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to claim 2, characterized in that, The learnable filter bank is implemented by a one-dimensional convolutional layer containing multiple parallel convolutional kernels, each corresponding to an FIR subband filter; each convolutional kernel parameter serves as the tap coefficient of the corresponding FIR subband filter, and the convolutional kernel parameters can be updated during training.

4. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to claim 3, characterized in that, Frequency regularization function of learnable filter banks Represented as: in, , They represent the first The preset passband frequency set and stopband frequency set of each FIR subband filter The preset positive stability constant; Indicates the first Tap coefficients of each FIR subband filter The frequency domain response obtained by performing a discrete Fourier transform. This represents a discrete frequency index variable.

5. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to claim 4, characterized in that, in, This represents the length of the FIR subband filter.

6. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to claim 5, characterized in that, Diversity regularization function Represented as: in, Indicates the first The, the The correlation matrix between the normalized convolution kernels of the FIR subband filters This indicates the total number of FIR subband filters.

7. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to claim 6, characterized in that, in, , They represent the first The, the The normalized convolution kernel parameters of each filter.

8. The UAV intelligent detection method based on opportunistic signals according to claim 7, characterized in that, For the Tap coefficients of each FIR subband filter Normalization is performed to obtain the first... Normalized convolution kernels for each FIR subband filter : in, It is a stable term.

9. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to any one of claims 1-8, characterized in that, The detection result of whether the drone exists is obtained by performing the following operations: The complex baseband sequence of real-time opportunity signals is input into the trained UAV intelligent detection model, and the multi-subband signal is predicted and output by the trained learnable filter bank. The multi-channel subband signal is input into the drone detection network, and the drone detection network predicts and outputs the detection result of whether the drone exists.

10. The intelligent detection method for unmanned aerial vehicles based on opportunistic signals according to claim 9, characterized in that, The opportunity signals include one or more of the following: FM broadcast signals, AM broadcast signals, digital audio broadcast signals, digital television broadcast signals, cellular communication downlink signals, wireless local area network signals, and satellite navigation signals.

Images