A method and system for separating signals of unmanned aerial vehicles based on strong interference adaptation

Abstract

The application discloses a kind of unmanned aerial vehicle signal separation method and system based on strong interference adaptation, the method is first obtained in target monitoring frequency band wideband unmanned aerial vehicle radio IQ data, generates wideband time-frequency diagram.Secondly, build including non-overlapping coexistence and overlapping coexistence inside and outside double cycle collaborative task pool, the wideband time-frequency diagram to be detected is input into trained inside and outside double cycle collaborative adaptive YOLO detector, and the rectangular detection frame of unmanned aerial vehicle signal and interference signal is output.Then according to rectangular detection frame, judge the overlapping relationship between unmanned aerial vehicle signal and interference signal, and construct signal overlapping relationship matrix.Finally, according to overlapping relationship matrix, select signal separation strategy, and output separated single unmanned aerial vehicle signal or interference signal.The application can simultaneously complete rectangular area detection of unmanned aerial vehicle signal and interference signal, overlapping relationship discrimination and scene separation, improve the adaptability to complex wideband unmanned aerial vehicle time-frequency diagram.

Description

Technical Field

[0001] This invention relates to the fields of radio signal monitoring, UAV signal detection, complex electromagnetic environment perception, time-frequency map target detection, signal recognition, signal separation, and intelligent signal processing technology, specifically to a UAV signal separation method and system based on strong interference adaptive design. Background Technology

[0002] In real-world broadband electromagnetic environments, multiple drone signals, Wi-Fi, Bluetooth, cellular communication signals, and other non-cooperative radiation sources often coexist within the same observation frequency band. Due to spectrum resource congestion and frequency reuse, different signals may overlap in the frequency domain, causing the time-frequency diagram obtained by the receiving equipment to exhibit phenomena such as signal boundary adhesion, overlapping energy regions, and interference covering drone signals, thus affecting the detection and identification of drone signals.

[0003] Existing UAV signal detection methods typically convert the received signal into a time-frequency map, and then use target detection networks or image classification networks to complete signal localization and identification. Since UAV image transmission signals and some OFDM signals usually appear as approximately rectangular regions on the time-frequency map, target detection networks can output the center coordinates, width, and height of the signal region, and further map them to the signal's center frequency, bandwidth, start and end times, and duration. Based on this detection box information, frequency band clipping, down-conversion, and low-pass filtering can be performed on the target signal, thereby achieving a certain degree of extraction of individual signals.

[0004] However, existing methods still have significant shortcomings: they lack explicit modeling of the overlapping relationships, temporal overlap, and time-frequency aliasing among multiple detection boxes; target detectors can only locate signal regions and cannot directly achieve strict separation of overlapping signals; fixed confidence thresholds and non-maximum thresholds are prone to accidentally deleting real signal boxes in overlapping scenarios, leading to missed detections, or introducing false detections and false alarms due to lowering the threshold.

[0005] Furthermore, the drone signal detection scenario is highly dynamic. Different drone models, communication protocols, modulation styles, signal-to-noise ratios, interference types, and degrees of overlap will cause the rectangular regions in the time-frequency plot to exhibit different scales, boundaries, energy distributions, and texture features. Fixed detector parameters, loss weights, post-processing thresholds, and separation parameters are difficult to adapt to multiple complex scenarios such as no overlap, frequency domain overlap, and time-frequency aliasing.

[0006] Therefore, an adaptive method and system for detecting and separating UAV signal overlap under strong interference is needed. This method should first use a target detection network to detect the UAV signal region and the interference signal region, then determine whether there is frequency domain overlap, temporal overlap, or time-frequency aliasing based on the detection box boundaries, and select the corresponding separation strategy according to different overlap types. The method should also be able to select the appropriate separation strategy based on the overlap type, using down-conversion and frequency domain separation based on detection box parameters for separable signals, and soft mask separation based on detection box constraints for time-frequency aliasing signals. Simultaneously, an internal and external dual-loop collaborative learning mechanism is introduced to adaptively adjust parameters during detection and separation, addressing issues such as insufficient adaptability of fixed parameters in complex electromagnetic environments, easy suppression of overlapping boxes, easy missed detection of weak signals, and unstable determination of overlap relationships. Summary of the Invention

[0007] To address the shortcomings of existing UAV signal detection methods based on time-frequency maps in scenarios where UAV signals and interference signals coexist, such as insufficient discrimination of detection box overlap, poor adaptability of fixed detection parameters, easy suppression of overlapping detection boxes, easy omission of weak UAV signals, and difficulty in effectively separating overlapping regions, this invention proposes an adaptive UAV signal overlap detection and separation method and system under strong interference.

[0008] This invention breaks through the limitations of traditional target detection methods that treat UAV signals as ordinary rectangular targets for separate detection. It constructs a cascaded processing architecture based on internal and external dual-loop collaborative adaptive detection, rectangular boundary overlap discrimination, and scene-specific signal separation. Through this cascaded mechanism, it achieves the detection, overlap discrimination, and separation of UAV signals and interference signals in broadband UAV time-frequency maps.

[0009] In one aspect, this invention provides a method for separating UAV signals based on strong interference adaptation, comprising the following steps:

[0010] Step 1: Acquire broadband UAV radio IQ data within the target monitoring frequency band, perform time-frequency transformation on it, and generate a broadband time-frequency map.

[0011] Step 2: Based on the UAV signals and interference signals in the wideband time-frequency map, construct an inner and outer dual-loop collaborative task pool including non-overlapping coexistence and overlapping coexistence, and train the inner and outer dual-loop collaborative adaptive YOLO detector.

[0012] Step 3: Input the broadband time-frequency graph to be detected into the trained dual-loop collaborative adaptive YOLO detector, and output the detection boxes of UAV signal and interference signal.

[0013] Step 4: Based on the detection frame output by the YOLO detector, determine the frequency domain overlap, time overlap, and time-frequency aliasing relationship between the UAV signal and the interference signal, and construct a signal overlap matrix.

[0014] Step 5: Select the corresponding signal separation strategy based on the overlap relationship matrix. For non-overlapping or separable signals, use frequency domain separation and clipping separation based on detection box parameters. For time-frequency aliasing signals, use soft mask separation based on detection box constraints. Finally, output the separated single UAV signal or interference signal.

[0015] Furthermore, in step 1, a short-time Fourier transform is performed on the broadband UAV radio IQ number to obtain the complex time spectrum of the mixed signal. Logarithmic enhancement and pseudo-color mapping are then applied to the amplitude of the complex time spectrum, and a broadband time-frequency diagram is generated based on the amplitude of the complex time spectrum. The horizontal axis of the broadband time-frequency diagram corresponds to the time dimension, and the vertical axis corresponds to the frequency dimension. The UAV signal and interference signal in the broadband time-frequency diagram are represented as rectangular regions with duration and bandwidth.

[0016] Furthermore, in step 2, in the task where the UAV signal and the interference signal coexist without overlap, the broadband time-frequency map sample simultaneously contains both the UAV signal and the interference signal. The two do not overlap in the frequency domain or the time-frequency domain. This is used to train the YOLO detector to simultaneously locate the UAV signal region and the interference signal region, and to learn the boundary differences between the two in a coexisting but separable scenario.

[0017] Furthermore, in step 2, in the task of overlapping and coexisting UAV signals and interference signals, the broadband time-frequency map sample contains both UAV signals and interference signals, and the two overlap in the frequency domain or time domain, or overlap in both the time and frequency domains, in order to train the YOLO detector to still output UAV signal frames and interference signal frames even when the signal regions overlap.

[0018] Furthermore, in step 2, independent bounding boxes are labeled for the UAV signal and interference signal in each broadband time-frequency map sample. For overlapping UAV signals and interference signals, the overlapping area is not labeled as a single target, but the complete bounding box labels of the UAV signal and the interference signal are retained separately, enabling the YOLO detector to learn the detection patterns of two independent signal instances in overlapping scenarios.

[0019] Furthermore, the aforementioned dual-loop collaborative adaptive YOLO detector is trained collaboratively through an inner loop and an outer loop. The inner loop enables the YOLO detector to adapt quickly within the same task type, while the outer loop is used for learning common parameters and optimizing parameter initialization across different task types. This allows the YOLO detector to adapt quickly with a small number of samples or gradient updates when facing scenarios where unknown UAV signals and interference signals overlap.

[0020] Furthermore, during the inner loop learning process, broadband time-frequency map samples are selected from the current task type to quickly update the network parameters, detection threshold, non-maximum suppression threshold, and separation parameters of the YOLO detector, so that the YOLO detector can adapt to the signal morphology, time-frequency boundary, energy distribution, and overlap characteristics under this task type.

[0021] Furthermore, the inner loop primarily learns the differences in local scenarios within the same task type. For tasks where UAV signals and interference signals coexist without overlap, the inner loop focuses on learning the detection box localization rules when the two signals are separable on the frequency and time axes. For tasks where UAV signals and interference signals coexist with overlap, the inner loop focuses on learning detection and separation parameters under scenarios such as signal overlap, boundary adhesion, strong interference covering weak UAV signals, and retention of highly overlapping candidate boxes.

[0022] Furthermore, during the outer loop learning process, broadband time-frequency map samples of different task types are used as evaluation criteria to perform cross-task performance evaluation on the YOLO detector after the inner loop update, and the shared initialization parameters and parameter adjustment strategies are updated based on the detection error, boundary localization error, overlap relationship discrimination error and separation error of each task.

[0023] Furthermore, the outer loop mainly learns the common patterns among different task types, including the rectangular structure of UAV signals and interference signals in the time-frequency graph, the changing patterns of detection box boundaries, the retention patterns of overlapping boxes, the discrimination patterns of overlapping relationships, and the adjustment patterns of separation parameters. This enables the system to quickly adapt to new coexisting or overlapping scenarios with only a small number of samples or iterations.

[0024] Furthermore, in step 3, each detection frame output by the YOLO detector includes a signal category, detection confidence level, center time coordinate, center frequency coordinate, duration width, and bandwidth height. The signal category includes at least drone signals and interference signals, and the detection confidence level represents the degree to which the YOLO detector believes that a category signal exists within the rectangle.

[0025] Furthermore, the detection frame output by the YOLO detector is converted into a time-frequency boundary form, including the signal start time, signal end time, lower frequency boundary, and upper frequency boundary. The signal start time and signal end time are used to determine the range of the signal on the time axis, while the lower and upper frequency boundaries are used to determine the range of the signal on the frequency axis.

[0026] Furthermore, in step 4, the existence of frequency overlap is determined based on the boundary relationship between the UAV signal detection frame and the interference signal detection frame on the frequency axis, and the existence of temporal overlap is determined based on their boundary relationship on the time axis. Additionally, the existence of time-frequency aliasing is determined based on the overlapping area of ​​the two frames on the time-frequency plane. If the UAV signal detection frame and the interference signal detection frame overlap on the frequency axis but not on the time axis, it is determined to be frequency-only overlap; if they overlap on the time axis but not on the frequency axis, it is determined to be temporal-only overlap; if they overlap on both the time and frequency axes, it is determined to be time-frequency aliasing; if they do not overlap on either the time or frequency axis, it is determined to be non-overlapping. The overlap relationship determination result between the UAV signal and the interference signal is then constructed.

[0027] Furthermore, in step 5, different signal separation strategies are selected based on the overlap relationship discrimination results. For signals without overlap, the target region is directly extracted based on the time and frequency boundaries corresponding to the detection box; for signals that only overlap in the time domain but not in the frequency domain, filters are designed based on the center frequencies and bandwidths of the UAV signal and the interference signal to perform frequency domain separation; for signals that only overlap in the frequency domain but not in time, time slicing is performed based on their respective start and end times; for time-frequency aliasing signals that overlap in both time and frequency, the soft mask separation process is initiated.

[0028] Furthermore, the frequency domain separation based on the detection frame parameters includes: down-converting the received UAV signal and interference signal according to the center frequency given by the detection frame, shifting the UAV signal and interference signal to the baseband or intermediate frequency position; setting the filter passband according to the bandwidth given by the detection frame to filter out other signals and noise outside the frequency band of the UAV signal and interference signal; and performing time clipping on the filtered signal according to the start and end times given by the detection frame to obtain the corresponding UAV signal or interference signal.

[0029] Furthermore, the soft mask separation process includes: determining the prior time-frequency ranges of the UAV signal detection frame and the interference signal detection frame output by the YOLO detector; generating a time soft window and a frequency soft window within the corresponding prior time-frequency range; generating a corresponding two-dimensional soft rectangle prior by orthogonally multiplying the time soft window and the frequency soft window; calculating a two-dimensional soft mask matrix using the two-dimensional soft rectangle prior; weighting the original complex time-frequency spectrum using the two-dimensional soft mask matrix to distribute the energy in the overlapping region according to the two instances of the UAV signal and the interference signal; and then performing an inverse time-frequency transformation on the separated complex time-frequency spectrum to obtain the separated UAV signal and the interference signal.

[0030] Furthermore, the soft mask does not rigidly assign each time-frequency unit to either the UAV signal or the interference signal, but instead assigns continuous weights to each time-frequency unit. For non-overlapping regions, the soft mask weights approximate the corresponding target signal; for overlapping regions, the soft mask adaptively allocates the energy contributions of the UAV signal and the interference signal based on the detection box prior, local energy distribution, boundary relationships, and separation parameters output by the inner and outer double-loop collaborative learning.

[0031] Furthermore, the soft mask separation parameters are adaptively adjusted by the inner and outer dual-loop collaborative learning modules according to the current task scenario. These parameters include overlapping region weight allocation coefficients and mask smoothing weights. When the overlap is light, the system employs weaker mask smoothing and narrower boundary expansion; when the overlap is heavy, the system enhances boundary expansion and overlapping region weight allocation capabilities to improve separation stability.

[0032] Furthermore, in step 5, the output results include the number of detected signals, the UAV signal detection frame, the interference signal detection frame, the signal type, the center frequency, the bandwidth, the start and end time, the detection confidence level, the overlap type, the overlap relationship discrimination result, and the separated UAV signal and interference signal.

[0033] In another aspect, the present invention also provides a UAV signal separation system based on strong interference adaptation, comprising the following modules:

[0034] The broadband time-frequency map module is used to acquire broadband UAV radio IQ data within the target monitoring frequency band, perform time-frequency transformation, and generate a broadband time-frequency map.

[0035] The YOLO detector module, based on UAV signals and interference signals in the wideband time-frequency map, constructs an inner and outer dual-loop collaborative task pool including non-overlapping coexistence and overlapping coexistence, and trains an inner and outer dual-loop collaborative adaptive YOLO detector.

[0036] The overlap matrix module is used to input the broadband time-frequency map to be detected into the trained dual-loop collaborative adaptive YOLO detector, and output rectangular detection boxes of UAV signals and interference signals. Based on the rectangular detection boxes output by the YOLO detector, the frequency domain overlap relationship, time overlap relationship and time-frequency aliasing relationship between UAV signals and interference signals are determined, and a signal overlap matrix is ​​constructed.

[0037] The UAV signal separation output module is used to select the corresponding signal separation strategy according to the overlap relationship matrix. For non-overlapping or separable signals, filtering and clipping separation based on detection box parameters are used. For time-frequency aliasing signals, soft mask separation based on detection box constraints is used. Finally, the separated single UAV signal or interference signal is output.

[0038] The beneficial effects of this invention include:

[0039] This invention proposes an adaptive method and system for detecting and separating overlapping UAV signals under strong interference, which can simultaneously perform rectangular region detection, overlap relationship discrimination, and scene-specific signal separation of UAV signals and interference signals.

[0040] This invention constructs an inner and outer dual-loop collaborative task pool consisting of tasks where UAV signals and interference signals coexist without overlap and tasks where UAV signals and interference signals coexist with overlap. This enables the YOLO detector to learn local differences within the same type of task and to learn transferable signal morphology and boundary features between different task types, thereby improving its adaptability to complex broadband UAV time-frequency maps.

[0041] This invention utilizes an internal and external dual-loop collaborative learning mechanism to adaptively adjust the confidence threshold, non-maximum suppression threshold, and loss function weights of the YOLO detector. This enables the YOLO detector to quickly adapt to the current scenario in the same type of task and maintain stable generalization ability across different task types, overcoming the problem that fixed detection parameters are difficult to adapt to different interference and overlapping scenarios.

[0042] This invention constructs an overlap relationship discrimination mechanism based on the time and frequency boundaries of the UAV signal detection frame and the interference signal detection frame. It can distinguish various cases such as no overlap, frequency domain overlap only, time domain overlap only, time-frequency aliasing, and inclusion overlap, providing a basis for subsequent separation strategy selection.

[0043] This invention employs different separation strategies for different overlap types. For signals that can be distinguished by frequency or time, filtering or cropping is used for separation. For signals that overlap in both time and frequency, soft masking is used for separation, thereby improving the adaptability of the separation method to complex scenarios. Attached Figure Description

[0044] Figure 1The overall flowchart for internal and external dual-loop adaptive signal overlap detection and separation;

[0045] Figure 2 The time-frequency diagram of the overlapping signal combination;

[0046] Figure 3 The time-frequency diagram is for combinations of non-overlapping signals;

[0047] Figure 4 A comparison of the Monte Carlo F1 scores of the dual-loop YOLO detector and the baseline YOLO detector over 100 iterations;

[0048] Figure 5 A comparison of the Monte Carlo accuracy of the dual-loop YOLO detector and the baseline YOLO detector over 100 trials;

[0049] Figure 6 A comparison of the Monte Carlo recall rates of the dual-loop YOLO detector and the baseline YOLO detector over 100 iterations;

[0050] Figure 7 The detection and separation results for instance mix_001469_tf_overlap_jsrp10dB_separation are shown in the image. Detailed Implementation

[0051] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below.

[0052] In one aspect, this invention provides a method for separating UAV signals based on strong interference adaptive signal separation, such as... Figure 1 As shown, it includes the following steps:

[0053] Step 1: Acquire broadband UAV IQ data and generate broadband time-frequency graph.

[0054] Specifically, acquire broadband UAV radio IQ data within the target monitoring frequency band. Under complex electromagnetic environments, the discrete broadband complex baseband signal model received within the current observation time window... It can be represented as:

[0055]

[0056] in, Index of discrete-time sampling points; Indicates the first One drone signal; This represents the total number of UAV signals simultaneously present within the observed frequency band. Indicates the first One interference signal; This represents the total number of interfering signals. This is the Gaussian white noise of the receiver system.

[0057] In order to analyze the time-frequency characteristics of the signal, the IQ data... A short-time Fourier transform is performed to obtain the complex time-frequency spectrum of the mixed signal. To suppress spectral leakage while maintaining time-frequency resolution, a Hamming window is chosen. Let the window function be... The window length is The number of FFT points is The time sliding step length is Then the complex time-frequency coefficients of the m-th frame and the k-th frequency point The calculation formula is:

[0058]

[0059] in, For time frame indexing, For discrete frequency indexing. The formula for the window function of the Hamming window is ( (for window length)

[0060]

[0061] Because deep learning networks are sensitive to numerical scales, directly using linear amplitude spectra can easily lead to strong interference signals masking weak UAV signals. Therefore, complex time-frequency spectra are used instead. The amplitude information is used to calculate the time-frequency plot, and then logarithmic enhancement is performed on it. The logarithmic power spectrum matrix is ​​calculated. The formula is:

[0062]

[0063] in, This is a very small normal value, and its function is to prevent... The logarithmic function is meaningless or produces negative infinity, thus ensuring the stability of numerical calculations.

[0064] To adapt to the standard image input requirements of subsequent YOLO detectors and convert the numerical matrix into a visual image form, it is necessary to globally normalize the logarithmic power spectrum and perform pseudo-color mapping.

[0065] First, use Min-Max normalization to... Mapped to the [0,1] interval, and then, using a pseudo-color mapping algorithm, the normalized two-dimensional matrix is... Converted to a three-channel broadband time-frequency diagram:

[0066]

[0067] in, This represents a pseudo-color mapping function. This is the generated broadband time-frequency map. This time-frequency map serves as the input feature matrix for the subsequent target YOLO detector.

[0068] Step 2: Construct an internal and external dual-loop collaborative task pool and train the YOLO detector.

[0069] First, a dual-loop collaborative task pool is constructed, which follows a specific task distribution and includes two types of sub-tasks: tasks where UAV signals and interference signals coexist without overlap. And tasks involving the coexistence of overlapping and interfering drone signals and jamming signals. The first task was sampled from the task pool. Specific tasks Each task's corresponding broadband time-frequency map sample set is strictly divided proportionally into a support set S and a query set Q. The support set is used for rapid adaptive updates of the inner loop, while the query set is used for global meta-parameter evaluation and updates of the outer loop. For all time-frequency map samples, independent bounding boxes are labeled for UAV signals and interference signals respectively. Even if time-frequency aliasing occurs, overlapping areas are not merged; instead, complete bounding box labels for the two independent signal instances are retained. The bounding box label format is defined as follows:

[0070]

[0071] Where c represents the signal category (drone or jamming signal). Centered on the time coordinate, Center frequency coordinates For duration width, This refers to bandwidth height.

[0072] For each task, the network parameters of the basic YOLO detector are set as follows: For complex overlapping fields

[0073] In this scenario, an adaptive weighted YOLO comprehensive loss function is constructed. For a given sample set Its loss function is defined as:

[0074]

[0075] in, For bounding box regression loss, For target confidence loss, For classification loss; , , This refers to the adaptive loss weights output by the dual-loop collaborative module.

[0076]

[0077] Where N is the number of samples, P is the sample set, and the j-th prediction box is... For the prediction box, This is the actual detection frame. This represents the cross-union ratio (CUI) between the predicted bounding box and the detected bounding box. The loss considers the overlap area, center distance, and aspect ratio difference between the predicted and ground truth bounding boxes.

[0078] set up Let be the total number of candidate positions, and let the existence of the true target at the j-th candidate position be denoted as . The confidence level of the predicted target is The target confidence loss is then expressed as:

[0079]

[0080] The inner loop's function is to enable the model to quickly adapt to the local signal morphology and overlapping features of the current time-frequency plot within the same task type. In the... Task In the inner loop training, the loss is calculated using the support set S, and gradient descent is used to quickly update the network parameters of the YOLO detector in one step. The parameter update formula is:

[0081]

[0082] in, For the inner loop learning rate, In order to adapt to the current situation after the internal loop update. Temporary network parameters for the task.

[0083] The outer loop primarily learns the common patterns among different task types (non-overlapping, time-frequency overlapping) and optimizes the global initial parameters. And update the hyperparameter tuning strategy.

[0084] The query set Q is fed into the temporary model parameters updated via the inner loop. In the middle, the cross-task validation loss is calculated, and the initial global parameters are adjusted accordingly. Perform an update. The gradient update formula for the outer loop is:

[0085]

[0086] in, The outer loop learning rate is used. While optimizing the outer loop, the inner and outer loop collaborative learning modules dynamically output adaptive adjustment parameters based on the high-dimensional feature representation of the current task.

[0087] Confidence threshold adjustment: When the feature map indicates the presence of weak UAV signals or strong noise background, the system outputs a dynamic confidence threshold. This threshold is lowered in weak signal scenarios to reduce missed detections, and raised in non-target areas with strong noise to suppress false alarms.

[0088] Step 3: Adaptive YOLO detection and time-frequency boundary transformation.

[0089] The time-frequency graph of the broadband UAV to be detected is input into the trained dual-loop collaborative adaptive YOLO detector. Adaptive threshold filtering and dynamic non-maximum suppression are used to output the effective detection box of the target signal and convert it into an absolute time-frequency boundary.

[0090] The preprocessed broadband UAV time-frequency map is input into the YOLO detector. The YOLO detector network performs forward propagation, outputting a series of initial detection boxes. For the j-th initial detection box, its parameterized vector is represented as:

[0091]

[0092] in, The signal type is either a drone signal or a jamming signal. To detect confidence level, which represents the degree to which the YOLO detector is confident that a class signal exists within the bounding box; and These are the center coordinates of the signal frame on the time axis and the center coordinates of the frequency axis, respectively. The duration width predicted by the signal; This represents the bandwidth height predicted for the signal.

[0093] Subsequently, for the j-th valid detection frame, its signal start time is calculated. Signal termination time Lower frequency boundary and frequency upper boundary :

[0094] ,

[0095] ,

[0096] The signal start and end times are used to determine the signal's occupancy on the time axis, while the lower and upper frequency boundaries are used to determine the signal's occupancy on the frequency axis. The final output is the target boundary matrix for the j-th signal. .

[0097] Step 4: Determining signal overlap based on time-frequency boundaries and constructing a matrix

[0098] Take all the high-precision boundary matrix sets output in step 3 and classify them according to the signal category. It is divided into a set of UAV signal targets and a set of interference signal targets.

[0099] Suppose that N drone signals and M interference signals are detected in the current time-frequency graph:

[0100] For the nth UAV signal, the time-frequency boundary parameter is extracted as follows: For the m-th interference signal, the extracted time-frequency boundary parameter is: .

[0101] Traverse all target pairs Calculate the one-dimensional absolute overlap length of the two axes on the time axis and frequency axis, respectively.

[0102] absolute overlap length of time The calculation formula is:

[0103]

[0104] absolute overlap length of time The calculation formula is:

[0105]

[0106] If the calculation result is 0, it indicates that there is no physical overlap in the corresponding dimension; if it is greater than 0, the result value is the specific time width or frequency bandwidth of the overlap.

[0107] To avoid "pixel-level pseudo-overlap" caused by blurred edges in the time-frequency map or slight fluctuations in the detection box, this embodiment introduces a tolerance threshold (time-domain tolerance). and frequency domain tolerance Based on the boundary relationship between the two on the time and frequency axes, a strict overlap state determination is performed, assuming the state function is... .

[0108] If the drone signal and the interference signal do not have effective overlap on either the time axis or the frequency axis, that is... and If so, it is determined that there is no overlap. .

[0109] If the two have a valid overlap on the time axis, but no overlap on the frequency axis, that is... and If so, it is determined to be a time-domain overlap only. If the two have a valid overlap on the frequency axis but no overlap on the time axis, that is... and If so, it is determined to be an overlap only in the frequency domain. If the two have effective overlap on both the time axis and the frequency axis, that is... and If so, it is determined to be time-frequency aliasing. .

[0110] Based on the above logical judgment results, the system constructs a global signal overlap matrix for the current broadband time-frequency graph. The matrix has dimensions of (i.e., number of drone signals × number of interference signals).

[0111] The element in the nth row and mth column of the matrix represents the overlap state between the nth UAV signal and the mth interference signal, expressed as:

[0112]

[0113] By constructing this overlap relationship matrix, the system performs structured digital encoding of the spatial topology and overlap relationship between all signals, thus providing a direct basis for the system to adaptively schedule differentiated separation strategies such as "time slicing", "frequency domain filtering" or "two-dimensional soft mask" according to different overlap types in step 5.

[0114] Step 5: Separate signals by scene.

[0115] For signals that do not undergo time-frequency two-dimensional aliasing, linear isolation is directly achieved in non-overlapping orthogonal dimensions using the physical parameters in the high-precision boundary matrix.

[0116] against Using the center frequency and bandwidth Construct a bandpass filter for broadband mixed signals. Perform digital downconversion and filtering operations:

[0117]

[0118] in, The cutoff frequency is The impulse response of the low-pass filter, This indicates a convolution operation.

[0119] against Utilizing time boundaries and Constructing a rectangular time window Perform time slicing:

[0120]

[0121] The output is separated as follows: .

[0122] against Based on the bounding box boundaries, a Sigmoid activation function is introduced. First, construct the soft window. Regarding the time dimension:

[0123] , ,

[0124] in , It is the time start boundary of the detection box. It is the time smoothing coefficient. Similarly, for the frequency dimension:

[0125] ,

[0126]

[0127] in , These are the upper and lower boundaries of the detection frame's frequency. It is the frequency smoothing coefficient.

[0128] Soft window of time and frequency soft window Orthogonal multiplication can generate a soft rectangular prior about the UAV signal. soft rectangular priors of interference signals :

[0129] ,

[0130] Under this prior model, the weights of the regions inside the detection box are close to 1, and the weights of the regions outside the detection box are close to 0, thus achieving a smooth transition at the boundaries.

[0131] Subsequently, a mask smoothing coefficient and overlapping region weight allocation coefficient, dynamically output by an inner and outer dual-loop collaborative learning module for the current aliasing task, are introduced. Calculate the two-dimensional soft mask matrix of the time-frequency unit. :

[0132]

[0133] in, To prevent extremely small constants with a denominator of zero, The spectrum when the number is complex. Coefficients It will make the boundaries sharper. This makes the energy transition smoother. Multiplying the two-dimensional soft mask matrix by the complex time-frequency spectrum of the original received mixed signal completes the adaptive weighted allocation of frequency domain energy in the aliasing region:

[0134]

[0135] For the separated complex time-frequency matrix and Perform inverse short-time Fourier transforms on each signal to recover the discrete-time signal sequence. and .

[0136] This method processes the broadband received data and outputs a comprehensive situational awareness and separation result of the UAV signal and interference signal. The output result specifically includes:

[0137] The total number of UAV signals and the total number of interference signals detected in the current broadband time-frequency graph;

[0138] Detailed parameters for each detection box include: signal type, detection confidence level, center time, center frequency, duration width, and bandwidth height;

[0139] The boundaries obtained from the conversion of each signal detection box include: signal start time, signal end time, lower frequency boundary, and upper frequency boundary;

[0140] The overlap type states between each signal pair include: no overlap, time-domain overlap only, frequency-domain overlap only, and time-frequency aliasing;

[0141] Complex time spectrum of the separated UAV signal and complex time spectrum of the interference signal.

[0142] Through the above steps, the present invention can simultaneously perform target detection, overlap relationship determination, and scene-specific signal separation of UAV signals and interference signals in a broadband UAV time-frequency map.

[0143] In another aspect, the present invention provides a UAV signal separation system based on strong interference adaptive design, for implementing the aforementioned UAV signal separation method, comprising a wideband time-frequency graph module, a YOLO detector module, an overlap relation matrix module, and a UAV signal separation output module:

[0144] The broadband time-frequency map module is used to acquire broadband UAV radio IQ data within the target monitoring frequency band, perform time-frequency transformation, and generate a broadband time-frequency map.

[0145] The YOLO detector module constructs an inner and outer dual-loop collaborative task pool, including non-overlapping coexistence and overlapping coexistence, based on the UAV signal and interference signal in the wideband time-frequency map, and trains the inner and outer dual-loop collaborative adaptive YOLO detector.

[0146] The overlap matrix module is used to input the broadband time-frequency map to be detected into the trained inner and outer dual-loop collaborative adaptive YOLO detector, and output rectangular detection boxes of UAV signals and interference signals; based on the rectangular detection boxes output by the YOLO detector, the frequency domain overlap relationship, time overlap relationship and time-frequency aliasing relationship between UAV signals and interference signals are determined, and a signal overlap matrix is ​​constructed.

[0147] The UAV signal separation output module is used to select the corresponding signal separation strategy according to the overlap relationship matrix. For non-overlapping or separable signals, filtering and clipping separation based on detection box parameters are used. For time-frequency aliasing signals, soft mask separation based on detection box constraints is used. Finally, the separated single UAV signal or interference signal is output.

[0148] In one specific embodiment, validation was performed based on a constructed dataset. The experimental environment consisted of a Windows 10 operating system, the PyTorch deep learning framework, and a single NVIDIA RTX 5060ti GPU.

[0149] Dataset Description: In this embodiment, a mixed signal dataset is constructed by selecting various UAV radio signals and artificially constructed broadband interference signals to verify the detection, overlap discrimination and separation capabilities of the present invention under the conditions of coexistence and overlap of UAV signals and interference signals.

[0150] The dataset used publicly available data included image transmission signal data from four DJI drone models: Mini 2, Mavic, Matrice, and Inspire 2.

[0151] The signal type is UAV image transmission signal (OFDM modulation), with a center frequency of 2.4 GHz. For each type of UAV signal, multiple signal segments are extracted as clean UAV signal samples. These UAV signal samples are then superimposed with interference signals to construct broadband hybrid signals with different overlap relationships.

[0152] Furthermore, the UAV IQ signal and the OFDM interference signal are linearly superimposed in the time domain to construct a hybrid signal dataset. The total number of samples in the hybrid signal dataset is 2000, of which 1600 samples are used for model training and 400 samples are used for validation and evaluation. Each hybrid signal sample contains one UAV signal component and one OFDM interference signal component, and the corresponding clean UAV signal component and clean interference signal component are simultaneously saved as reference ground truth values ​​for signal separation performance evaluation. The interference-to-signal ratio (JSR) covers five levels: -10dB, -5dB, 0dB, +5dB, and +10dB, with a step size of 5dB.

[0153] For each hybrid IQ signal, a complex time spectrum is generated using a short-time Fourier transform (SFT). In this embodiment, the SFT uses a Hamming window with a window length of 2048, an overlap length of 1024, and 2048 FFT points. The power spectrum is calculated from the complex time spectrum and logarithmic enhancement is performed to improve the visibility of weak UAV signals and signal boundary regions.

[0154] Furthermore, the enhanced logarithmic power spectrum is normalized to map its values ​​to the [0,1] interval; subsequently, a pseudo-color mapping method is used to convert the two-dimensional time-frequency matrix into a three-channel RGB time-frequency image, with the image size set to 640×640. This time-frequency image serves as the input to the YOLO detector, such as... Figure 2 , Figure 3 . Figure 2 The diagram shows a separate UAV signal and a separate OFDM signal, as well as a time-frequency diagram showing the aliasing of the two in the time-frequency domain. Figure 3 The display shows individual drone signals and individual OFDM signals, as well as time-frequency diagrams showing that the two do not overlap.

[0155] For each time-frequency map sample, separate bounding boxes are labeled for the UAV signal and the interference signal. Even if time-frequency overlap occurs, the overlapping area is not labeled as a single target; instead, the complete detection boxes for the UAV signal and the interference signal are preserved separately. Labels are in YOLO format and are represented as follows: , , , , ,in, Indicates the signal type: 0 indicates a drone signal, and 1 indicates a jamming signal; and These represent the normalized coordinates of the detection frame center on the time axis and frequency axis, respectively; width and height represent the normalized values ​​of the signal duration width and bandwidth height, respectively.

[0156] This embodiment first constructs a baseline YOLO detector. The baseline model adopts the YOLOv5s architecture, with the backbone network being CSPDarknet53, the neck network using an FPN+PAN feature pyramid structure, and the detection head using a standard Detect head. The model is initialized by loading a weight file pre-trained on the COCO dataset, and the number of output categories is modified to 2 (drone signals and interference signals).

[0157] Based on the baseline detector model, an inner loop is performed for scene-by-scene fine-tuning. The goal of the inner loop is to enable the detection model to be fine-tuned specifically for non-overlapping scenes and time-frequency overlapping scenes, based on the baseline model, and to learn the unique local signal feature differences and boundary morphology of each scene.

[0158] For non-overlapping scenarios, spectrogram-based optimization of the hyperparameter file is used for fine-tuning. Key adjustments to this hyperparameter configuration include: disabling data augmentation, disabling color space enhancement, disabling vertical flipping, and retaining horizontal flipping; setting the initial target loss weight obj to 1.5, with the inner and outer double loop modules generating loss weight correction coefficients based on task characteristics; and switching the optimizer to the Adam adaptive optimizer. The fine-tuning rounds are set to 5 epochs, and the fine-tuning strategy involves updating only the parameters of the detector head and the last few layers of the backbone network, while freezing the shallow general feature extraction layers.

[0159] For time-frequency overlapping scenarios, a high-recall hyperparameter configuration was used for fine-tuning. This hyperparameter configuration further reduced the IoU training threshold (iou_t=0.15) based on spectrogram optimization to encourage the model to generate effective candidate detection boxes even under conditions of signal boundary adhesion and strong background interference. The initial target loss weight was also set to obj=1.5, and the inner and outer loop modules generated loss weight correction coefficients based on task characteristics. The optimizer used was Adam, and the fine-tuning rounds were 5 epochs. Through inner loop fine-tuning, the model mastered the ability to accurately locate signals in scenarios where the frequency and time domains are separable, and the ability to distinguish and detect overlapping signal entities in time-frequency aliasing scenarios.

[0160] After fine-tuning the inner loop, an outer loop is executed for cross-scene adaptive threshold search. The goal of the outer loop is to perform global coordination across different scene types, searching for the confidence threshold (conf) and NMS intersection-over-union threshold (iou) that optimize the overall detection performance for each scene.

[0161] The threshold search space is set as follows: confidence threshold conf∈{0.10, 0.20, 0.30, 0.40}, NMS intersection-union ratio threshold iou∈{0.30, 0.45, 0.60}, constituting 12 candidate parameter combinations. For each parameter combination, evaluation is performed on validation subsets of non-overlapping and time-frequency overlapping scenarios, with the F1 score as the optimization objective. The search process iterates for two rounds, using an exponential moving average (EMA) strategy to smoothly update the thresholds, with a smoothing coefficient... .

[0162] The final adaptive parameters obtained from the outer loop search are: the optimal thresholds for non-overlapping scenarios are conf=0.20 and iou=0.40; the optimal thresholds for time-frequency overlapping scenarios are conf=0.15 and iou=0.32. This result reveals a key technical finding: in time-frequency overlapping scenarios, the YOLO detector needs to use an adaptive confidence threshold (0.15) to reduce the risk of missed detections due to signal boundary aliasing, and simultaneously needs to use an adaptive NMS intersection-union threshold (0.32) to avoid incorrectly suppressing two overlapping but different category valid detection boxes. This scenario-differentiated threshold configuration verifies the necessity and effectiveness of the outer loop cross-scenario coordination mechanism.

[0163] During the inference phase, the dual-loop collaborative adaptive YOLO detector operates as follows: First, based on the time-frequency boundary obtained from the initial YOLO detection frame transformation, the time overlap length and frequency overlap length between the UAV signal and the interference signal are calculated, and the overlapping scene type to which the current time-frequency map belongs is determined accordingly. Then, the inner loop fine-tuning model weights corresponding to the scene are automatically selected and loaded; simultaneously, the optimal adaptive threshold parameters (conf, iou) for the scene obtained from the outer loop search are automatically applied.

[0164] During detection inference, the input image is first preprocessed using letterboxes to maintain the original aspect ratio and padded to 640×640 pixels. Then, it undergoes forward inference through the YOLOv5s network, outputting a series of candidate detection boxes. After NMS post-processing for deduplication, each valid detection box outputs a six-tuple [x1, y1, x2, y2, confidence, ...]. ], where (x1, y1) and (x2, y2) are the pixel coordinates of the top-left and bottom-right corners of the detection box in the 640×640 pixel image coordinate system, respectively, and confidence∈[0,1] is the detection confidence. ∈{0,1} is the signal category identifier.

[0165] The pixel coordinate detection box output by the YOLO detector is converted into a physical time-frequency bounding box. Furthermore, an N*M dimensional overlap matrix is ​​constructed to systematically characterize the time-frequency overlap state between all UAV-interference signal pairs.

[0166] Based on the determined overlap type, the corresponding signal separation strategy is adaptively selected.

[0167] The evaluation was performed on a validation set of 400 images. The detection performance metrics for each scenario are shown in Table 1 below.

[0168] Table 1

[0169]

[0170] As shown in the table above, the YOLO detector achieved perfect detection performance in all three scenarios: non-overlapping, time-domain overlapping, and frequency-domain overlapping. In the most challenging time-frequency aliasing scenario, the precision was 0.990, the recall was 0.971, and the F1 score was 0.980. These results demonstrate that the detection method of this invention exhibits excellent detection performance in all types of overlapping scenarios.

[0171] To verify the effectiveness of the dual-loop collaborative training mechanism, the dual-loop model and the baseline model were compared and evaluated under different scenarios. The parameters are shown in Table 2 below, and the results of 100 Monte Carlo experiments are presented as follows. Figure 4 , Figure 5 , Figure 6 .

[0172] Table 2

[0173]

[0174] The above comparison results demonstrate that the dual-loop training model significantly improves performance compared to the baseline detector model across all scenarios. This result strongly validates the effectiveness of the dual-loop collaborative training mechanism—the inner loop fine-tuning enables the model to grasp the local feature differences across scenarios, while the outer loop adaptive threshold search provides differentiated optimal inference parameters for different scenarios. The synergistic effect of both achieves a comprehensive improvement in detection performance.

[0175] Table 3

[0176]

[0177] As shown in Table 3 above, this scheme exhibits excellent separation performance in all four scenarios: non-overlapping, time-domain overlapping, frequency-domain overlapping, and time-frequency overlapping.

[0178] Taking the mixed signal sample mix_001469 (scene type: tf_overlap time-frequency aliasing, interference-to-signal ratio JSR=+10dB) as an example, the final output of the complete pipeline of this patent is illustrated. The interference signal power of this sample is 10 times that of the UAV signal power, belonging to the most extreme scenario of strong interference and full time-frequency overlap.

[0179] (1) YOLO detection results: Two targets were detected. The detection confidence level of the UAV signal was 0.983, and its time-frequency boundaries were: start time 0ms, end time 2ms, lower frequency limit -15.28MHz, upper frequency limit 2.92MHz, center frequency -6.18MHz, and bandwidth 18.2MHz. The detection confidence level of the OFDM interference signal was 0.923, and its time-frequency boundaries were: start time 0ms, end time 2ms, lower frequency limit -9.10MHz, upper frequency limit 9.10MHz, center frequency 0.00MHz, and bandwidth 18.20MHz.

[0180] (2) Overlap determination: The overlap between the UAV signal and the interference signal in the frequency dimension is: The amount of overlap in the time dimension is Therefore, it is determined to be a time-frequency aliasing type (overlap type code = 3).

[0181] (3) Separation Strategy Selection and Execution: Due to the determination of time-frequency aliasing, the system automatically triggers the third separation strategy—STFT two-dimensional soft mask separation. During the separation process, a complex STFT matrix is ​​constructed, and soft masks for the UAV signal and the interference signal are generated on this matrix. The power ratio adaptive weight of the overlapping area is dynamically allocated according to the local power distribution of the two signals. After the separation is completed, the time-frequency diagrams of UAV signal separation and interference signal separation, as well as the corresponding complex time-frequency matrix data, are generated.

[0182] (4) Separation Quality Assessment: The SINR of the separated UAV signal was -5.2dB, and the SINR of the separated interference signal was -1.7dB. In this sample, the power of the interference signal was 10 times that of the UAV signal (JSR = +10dB), and the strong interference completely covered the time-frequency region of the UAV signal. The signal-to-interference ratio (SINR) in the original mixed signal was -10dB. After separation processing, the SINR of the UAV signal was improved to -5.2dB, achieving an effective SINR improvement of approximately 4.8dB. Considering that this sample belongs to the most difficult extreme scenario of time-frequency aliasing superimposed with 10 times the power of strong interference, the above separation results verify the effectiveness of the Sigmoid soft mask adaptive separation strategy proposed in this invention under extreme conditions.

[0183] (5) Comprehensive Output: The pipeline ultimately outputs a structured detection report in JSON format, which includes: the total number of detected UAV signals (1) and the total number of interfering signals (1); complete parameters for each detection box (signal type, confidence level, center time and center frequency, time width and bandwidth, start and end times, lower and upper frequency bounds); overlap matrix (1 pair of signals, overlap type is time-frequency aliasing); and the separated time-frequency plot as shown below. Figure 7 , Figure 7The first row and first column indicate the overlap between the UAV signal and the OFDM signal. The green detection box represents the UAV signal, and the yellow detection box represents the OFDM interference signal. The first row and second column represent the separated UAV signal, and the first row and third column represent the original UAV signal. The second row and first column represent the aliasing state when there is no detection box. The second row and second column represent the separated OFDM interference signal, and the second row and third column represent the clean OFDM interference signal. The third row and first column represent the UAV signal extraction area, and the third row and second column represent the OFDM interference signal extraction area. The yellow part represents the clean area without aliasing, and the blue part represents the area for weight allocation. The third row and third column represent the overall extraction area.

[0184] The above experimental results were obtained on the simulated mixed dataset and the same distribution validation set constructed in this embodiment, which are used to illustrate the effectiveness of the method of the present invention in typical UAV signal and OFDM interference coexisting scenarios.

Claims

1. A method for separating UAV signals based on strong interference adaptive behavior, characterized in that, Includes the following steps: Step 1: Acquire broadband UAV radio IQ data within the target monitoring frequency band, perform time-frequency transformation, and generate a broadband time-frequency map; Step 2: Based on the UAV signals and interference signals in the wideband time-frequency map, construct an inner and outer dual-loop collaborative task pool including non-overlapping coexistence and overlapping coexistence, and train the inner and outer dual-loop collaborative adaptive YOLO detector. Step 3: Input the broadband time-frequency map to be detected into the trained dual-loop collaborative adaptive YOLO detector, and output the rectangular detection boxes of UAV signal and interference signal; Step 4: Determine the frequency domain overlap, time overlap, and time-frequency aliasing relationship between the UAV signal and the interference signal based on the rectangular detection box output by the YOLO detector, and construct a signal overlap matrix; Step 5: Select the corresponding signal separation strategy based on the overlap relationship matrix. For non-overlapping or separable signals, use filtering and clipping separation based on detection box parameters. For time-frequency aliasing signals, use soft mask separation based on detection box constraints. Finally, output the separated single UAV signal or interference signal.

2. The UAV signal separation method based on strong interference adaptive as described in claim 1, characterized in that, In step 1, a short-time Fourier transform is performed on the broadband UAV radio IQ number to obtain the complex time spectrum of the mixed signal. Logarithmic enhancement and pseudo-color mapping are then applied to the amplitude of the complex time spectrum. A broadband time-frequency diagram is generated based on the amplitude of the complex time spectrum. The horizontal axis of the broadband time-frequency diagram corresponds to the time dimension, and the vertical axis corresponds to the frequency dimension. The UAV signal and interference signal in the broadband time-frequency diagram are represented as rectangular areas with duration and bandwidth.

3. The UAV signal separation method based on strong interference adaptive as described in claim 2, characterized in that, In step 2, in the task of non-overlapping coexistence of UAV signals and interference signals, the broadband time-frequency map sample contains both UAV signals and interference signals. The two do not overlap in the frequency domain or the time-frequency domain. This is used to train the YOLO detector to simultaneously locate the UAV signal region and the interference signal region, and to learn the boundary differences between the two in the coexisting but separable scenario. In tasks where UAV signals and interference signals coexist, broadband time-frequency map samples contain both UAV signals and interference signals, and these signals overlap in the frequency or time domain, or both in the time and frequency domains. This is used to train the YOLO detector to still output both UAV signal frames and interference signal frames even when signal regions overlap.

4. The UAV signal separation method based on strong interference adaptive as described in claim 3, characterized in that, In step 2, independent bounding boxes are labeled for the UAV signal and interference signal in each broadband time-frequency map sample. For overlapping UAV signals and interference signals, the overlapping area is not labeled as a single target. Instead, the complete bounding box labels of the UAV signal and the interference signal are retained respectively, so that the YOLO detector can learn the detection rules of two independent signal instances in overlapping scenarios.

5. The UAV signal separation method based on strong interference adaptive as described in claim 4, characterized in that, The inner and outer dual-loop collaborative adaptive YOLO detector is trained collaboratively through an inner loop and an outer loop, as specifically implemented as follows: During the inner loop learning process, broadband time-frequency map samples are selected from the current same task type to update the network parameters, detection threshold, non-maximum suppression threshold, and separation parameters of the YOLO detector. The inner loop learns the local scene differences under the same task type. For tasks where UAV signals and interference signals do not overlap and coexist, the inner loop learns the detection box localization rules when the two can be separated on the frequency axis and time axis. For tasks where UAV signals and interference signals overlap and coexist, the inner loop focuses on learning the detection and separation parameters under scenarios of signal overlap, boundary adhesion, strong interference covering weak UAV signals, and overlapping candidate box retention. During the outer loop learning process, broadband time-frequency map samples of different task types are used as evaluation criteria to evaluate the cross-task performance of the YOLO detector after the inner loop update. The shared initialization parameters and parameter adjustment strategies are updated based on the detection error, boundary localization error, overlap relationship discrimination error, and separation error of each task. The outer loop learns the common patterns among different task types, including the rectangular structure of UAV signals and interference signals in the time-frequency map, the change pattern of detection box boundaries, the pattern of overlapping box retention, the pattern of overlap relationship discrimination, and the pattern of separation parameter adjustment.

6. The UAV signal separation method based on strong interference adaptive as described in claim 5, characterized in that, In step 3, each rectangular detection box output by the YOLO detector includes signal category, detection confidence, center time coordinate, center frequency coordinate, duration width, and bandwidth height. The rectangular detection box output by the YOLO detector is converted into a time-frequency boundary form, including the signal start time, signal end time, lower frequency boundary, and upper frequency boundary. The signal start time and signal end time are used to determine the range occupied by the signal on the time axis, and the lower frequency boundary and upper frequency boundary are used to determine the range occupied by the signal on the frequency axis.

7. The UAV signal separation method based on strong interference adaptive as described in claim 6, characterized in that, In step 4, the existence of frequency overlap is determined based on the boundary relationship between the UAV signal detection frame and the interference signal detection frame on the frequency axis, the existence of time overlap is determined based on the boundary relationship between the two on the time axis, and the existence of time-frequency aliasing is determined based on the overlapping area of ​​the two on the time-frequency plane. If the UAV signal detection frame and the interference signal detection frame overlap on the frequency axis but not on the time axis, it is determined to be frequency domain overlap only; if they overlap on the time axis but not on the frequency axis, it is determined to be time domain overlap only; if they overlap on both the time axis and the frequency axis, it is determined to be time-frequency aliasing; if they do not overlap on either the time axis or the frequency axis, it is determined to be non-overlapping, and the overlap relationship discrimination result between the UAV signal and the interference signal is constructed.

8. The UAV signal separation method based on strong interference adaptive as described in claim 7, characterized in that, Step 5 is specifically implemented as follows: for signals without overlap, the target region is extracted directly based on the time and frequency boundaries corresponding to the detection frame; for signals that only overlap in the time domain but not in the frequency domain, filters are designed based on the center frequency and bandwidth of the UAV signal and the interference signal to perform frequency domain separation; for signals that only overlap in the frequency domain but not in time, time slices are performed based on their respective start and end times; for time-frequency aliasing signals that overlap in both time and frequency, the soft mask separation process is initiated.

9. The UAV signal separation method based on strong interference adaptive as described in claim 8, characterized in that, The frequency domain separation includes: downconverting the received UAV signal and interference signal to the baseband or intermediate frequency position according to the center frequency given by the detection frame; setting the filter passband according to the bandwidth given by the detection frame to filter out other signals and noise outside the frequency band of the UAV signal and interference signal; and time-trimming the filtered signal according to the start and end times given by the detection frame to obtain the corresponding UAV signal or interference signal. The soft mask separation process includes: determining the prior time-frequency ranges of the UAV signal detection frame and the interference signal detection frame output by the YOLO detector; generating a time soft window and a frequency soft window within the corresponding prior time-frequency range; generating a corresponding two-dimensional soft rectangle prior by orthogonally multiplying the time soft window and the frequency soft window; calculating a two-dimensional soft mask matrix using the two-dimensional soft rectangle prior; weighting the original complex time-frequency spectrum using the two-dimensional soft mask matrix to distribute the energy in the overlapping region according to the two instances of the UAV signal and the interference signal; and then performing an inverse time-frequency transformation on the separated complex time-frequency spectrum to obtain the separated UAV signal and the interference signal.

10. A UAV signal separation system based on strong interference adaptive design, used to implement the UAV signal separation method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The broadband time-frequency map module is used to acquire broadband UAV radio IQ data within the target monitoring frequency band, perform time-frequency transformation, and generate a broadband time-frequency map. The YOLO detector module, based on UAV signals and interference signals in the wideband time-frequency map, constructs an inner and outer dual-loop collaborative task pool including non-overlapping coexistence and overlapping coexistence, and trains an inner and outer dual-loop collaborative adaptive YOLO detector. The overlap matrix module is used to input the broadband time-frequency map to be detected into the trained inner and outer dual-loop collaborative adaptive YOLO detector, and output rectangular detection boxes of UAV signals and interference signals; based on the rectangular detection boxes output by the YOLO detector, the frequency domain overlap relationship, time overlap relationship and time-frequency aliasing relationship between UAV signals and interference signals are determined, and a signal overlap matrix is ​​constructed. The UAV signal separation output module is used to select the corresponding signal separation strategy according to the overlap relationship matrix. For non-overlapping or separable signals, filtering and clipping separation based on detection box parameters are used. For time-frequency aliasing signals, soft mask separation based on detection box constraints is used. Finally, the separated single UAV signal or interference signal is output.

Images