Unmanned aerial vehicle frequency hopping signal detection method and system based on clustering algorithm

Abstract

The invention discloses an unmanned aerial vehicle frequency hopping signal detection method and system based on a clustering algorithm, and belongs to the technical field of unmanned aerial vehicle measurement, and the method comprises the steps: carrying out the short-time Fourier transform of an electromagnetic signal, obtaining frequency spectrum data, carrying out the noise filtering of the frequency spectrum data, and obtaining a frequency hopping signal point; the frequency hopping signal points are clustered to generate historical observation clusters, two groups of parent hypotheses are created for the historical observation clusters of the same type, and each group of parent hypotheses comprises the number of unmanned aerial vehicles, hypothesis confidence and hypothesis trajectories; generating a current observation cluster based on the re-collected electromagnetic signals, and distributing the current observation cluster to the parent hypothesis to obtain child hypotheses; and calculating the assumption confidence of each sub-assumption, filtering the sub-assumptions based on the assumption confidence, and taking the sub-assumptions as final detection results of the frequency hopping signals when the assumption confidence of the sub-assumptions is higher than that of all other sub-assumptions in a plurality of continuous processing periods. According to the invention, the accuracy of unmanned aerial vehicle number detection can be improved.

Description

Technical Field

[0001] This invention belongs to the field of UAV monitoring technology, specifically relating to a method and system for detecting UAV frequency hopping signals based on clustering algorithms. Background Technology

[0002] In recent years, drone technology has developed rapidly and has been widely used in both civilian and military fields. However, the widespread adoption of drone technology has also brought a series of security risks. With the continuous development of drone technology, low-altitude security defense has become an important issue. In a low-altitude defense system, the effective detection, identification, and tracking of drones is the primary link and core challenge. Detecting the communication and control signals of drones through radio spectrum monitoring technology is a key technical means to achieve this goal.

[0003] To effectively detect UAV signals, various methods have been proposed in the prior art. For example, Chinese patent document CN110334591A discloses a method for detecting and identifying UAV frequency-hopping signals based on cluster analysis. This method first preprocesses the collected spectrum data to extract effective signals. Then, it performs cluster analysis on the extracted frequency-hopping signal points according to multiple dimensions such as signal bandwidth, waveform characteristics, energy, and dwell time. The clustering results are then sorted according to the principle of non-overlapping time to separate the signals of different UAVs. Finally, identification is achieved by comparing with a reference database. This method has achieved the detection of UAV frequency-hopping signals to a certain extent.

[0004] However, while the aforementioned method can separate UAV signals from mixed signals, it assumes that frequency-hopping signals generated by a single type of UAV will not exhibit two or more signal characteristics at the same time. This method relies excessively on the physical characteristics of the signals and simple temporal exclusivity. When multiple UAVs of the same model operate in the same airspace, their signals are highly similar in physical characteristics such as bandwidth, waveform, energy, and dwell time. This can lead to the method incorrectly classifying them into one category during the initial clustering stage. Subsequently, if the frequency-hopping sequences of multiple UAVs do not overlap temporally within the observation time, the method will misclassify these multiple independent UAVs as one, resulting in serious misdetection of the number of targets. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for detecting frequency hopping signals of unmanned aerial vehicles (UAVs) based on clustering algorithms, thereby resolving the issues present in the background art.

[0006] To achieve the aforementioned objectives, this invention proposes a method for detecting UAV frequency hopping signals based on clustering algorithms, comprising: Electromagnetic signals of a preset duration are collected, and short-time Fourier transform is performed on the electromagnetic signals to obtain spectral data. Noise is filtered from the spectral data to obtain frequency hopping signal points. The frequency hopping signal points are clustered to generate historical observation clusters. For the same type of historical observation cluster, two sets of parent hypotheses are created. Each set of parent hypotheses includes the number of UAVs, the hypothesis confidence level, and the hypothesis trajectory. The hypothesis trajectory includes the sequence formed by the historical observation clusters corresponding to each UAV in the parent hypotheses. The current observation cluster is generated based on the electromagnetic signals acquired again, and the current observation cluster is assigned to the parent hypothesis in different ways to obtain multiple sets of sub-hypotheses; Calculate the goodness of fit between the current observation cluster and the historical observation cluster for each sub-hypothesis, update the hypothesis confidence of the sub-hypothesis based on the goodness of fit, filter the sub-hypotheses based on the updated hypothesis confidence, and keep the sub-hypotheses for the next processing cycle. If the confidence level of a sub-hypothesis is higher than that of all other sub-hypotheses for multiple consecutive processing cycles, then that sub-hypothesis is taken as the final detection result of the frequency hopping signal.

[0007] Furthermore, clustering the frequency-hopping signal points to generate historical observation clusters includes the following steps: Time-frequency distance statistics are performed on the frequency hopping signal points and a distance distribution histogram is generated. A candidate neighborhood radius set is generated based on the peak position of the distance distribution histogram. The candidate neighborhood radius set includes multiple neighborhood radius parameters. DBSCAN clustering is performed based on the candidate neighborhood radius set to obtain multiple clustering results. A consensus evidence matrix is ​​constructed based on the number of times any two frequency-hopping signal points are classified into the same cluster in all clustering results. The frequency-hopping signal points are used as nodes to construct a graph structure. The edge weights between nodes in the graph structure are determined based on the consensus evidence matrix. After removing edges with edge weights less than a preset voting threshold, a corrected graph is obtained. Connectivity analysis is performed on the corrected graph, and nodes with connectivity are used as initial observation clusters. Edge correction is performed on the initial observation clusters to obtain historical observation clusters.

[0008] Furthermore, edge correction of the initial observation cluster includes the following steps: The frequency hopping signal points that have not been clustered are taken as target points. A fusion neighborhood is generated based on the target points. The frequency hopping signal points that have been clustered in the fusion neighborhood are taken as belonging points. If the proportion of belonging points belonging to the same initial observation cluster exceeds the belonging threshold, the target points are merged into the initial observation cluster. This step is repeated until each target point is traversed to complete the edge correction of the initial observation cluster.

[0009] Furthermore, creating two sets of parent hypotheses for the same type of historical observation clusters includes the following steps: The frequency and time of centroid points are extracted from historical observation clusters in chronological order to form a frequency time series. The frequency time series is reconstructed in phase space to obtain the state vector of each historical observation cluster. The historical observation clusters are clustered based on the distance between the state vectors to obtain UAV clusters. The number of UAV clusters is determined as the number of UAVs K. The separation degree between UAV clusters is calculated. The average value of all separation degrees is mapped to the first confidence degree corresponding to the number of UAVs K. Calculate the state transition probability between drone clusters, determine the transition path based on the state transition probability, correct the number of drones K based on the number of clusters included in the transition path, obtain the number of drones I, and calculate the second confidence level corresponding to the number of drones I based on the average state transition probability in the transition path. Generate hypothetical trajectories corresponding to the number of drones K and the number of drones I. Use the number of drones K, the first confidence level, and the corresponding hypothetical trajectory as the first set of parent hypotheses, and use the number of drones I, the second confidence level, and the corresponding hypothetical trajectory as the second set of parent hypotheses.

[0010] Furthermore, calculating the second confidence level corresponding to the number of drones I based on the average state transition probability in the transition path includes the following steps: The UAV clusters included in the transition path are taken as target clusters. The state transition probabilities between target clusters are organized into a numerical set. The average value of the numerical set is calculated to obtain the second confidence level.

[0011] Furthermore, generating the hypothetical trajectories corresponding to the number of drones K and I includes the following steps: Obtain the historical observation clusters included in each UAV cluster in the parent hypothesis, arrange the historical observation clusters in chronological order to obtain the trajectory groups of the corresponding UAV clusters, allocate and integrate the trajectory groups to obtain the hypothetical trajectories corresponding to the number of UAVs K and the number of UAVs I.

[0012] Further, calculating the goodness of fit between the current observation cluster and the historical observation cluster for each sub-hypothesis, and updating the hypothesis confidence of the sub-hypothesis based on the goodness of fit includes the following steps: Calculate the time interval sequence and frequency interval sequence of historical observation clusters in the parent hypothesis. Calculate the actual time interval and actual frequency interval between the current observation cluster and each historical observation cluster in each hypothesis trajectory. Count the number of occurrences of each actual time interval and actual frequency interval in the time interval sequence and frequency interval sequence. Select the highest occurrence count as the time interval occurrence count and frequency interval occurrence count. Calculate the goodness of fit based on the time interval occurrence count and frequency interval occurrence count. Combine the goodness of fit with the parent hypothesis using weighted fusion to obtain the updated confidence score. Normalize the updated confidence scores of all sub-hypotheses to obtain the hypothesis confidence scores of the sub-hypotheses.

[0013] This invention also provides a UAV frequency hopping signal detection system based on a clustering algorithm. This system is used to implement the methods described above, and includes: The filtering module acquires electromagnetic signals for a preset duration, performs a short-time Fourier transform on the electromagnetic signals to obtain spectral data, and performs noise filtering on the spectral data to obtain frequency hopping signal points. The first analysis module clusters the frequency hopping signal points to generate historical observation clusters. For the same type of historical observation cluster, two sets of parent hypotheses are created. Each set of parent hypotheses includes the number of UAVs, the hypothesis confidence level, and the hypothesis trajectory. The hypothesis trajectory includes the sequence formed by the historical observation clusters corresponding to each UAV in the parent hypotheses. The second analysis module generates the current observation cluster based on the re-acquired electromagnetic signals, assigns the current observation cluster to the parent hypothesis in different ways to obtain multiple sets of sub-hypotheses, calculates the fitting degree between the current observation cluster and the historical observation cluster in each sub-hypothesis, updates the hypothesis confidence of the sub-hypothesis based on the fitting degree, filters the sub-hypotheses based on the updated hypothesis confidence, and retains the sub-hypotheses to enter the next processing cycle. The output module takes a sub-hypothesis as the final detection result of the frequency hopping signal if the hypothesis confidence of the sub-hypothesis is higher than that of all other sub-hypotheses for multiple consecutive processing cycles.

[0014] The beneficial effects of this invention are as follows: This invention clusters frequency-hopping signal points to generate historical observation clusters. For the same type of historical observation cluster, two sets of parent hypotheses are created, each containing a different number of drones. Each parent hypothesis includes the number of drones, hypothesis confidence, and hypothesis trajectory, thus providing a possibility for correctly identifying the number of targets. Then, the current observation cluster is generated based on the re-acquired electromagnetic signals, and the current observation cluster is assigned to the parent hypotheses in different ways to obtain multiple sets of sub-hypotheses. The hypothesis confidence of the sub-hypotheses is dynamically updated, thereby continuously tracking the probability of multiple parent hypotheses and avoiding the high risk of a single judgment. Finally, only when the hypothesis confidence of a certain sub-hypothesis is consistently higher than all other sub-hypotheses for multiple consecutive processing cycles is that sub-hypothesis used as the final detection result of the frequency-hopping signal, ensuring the reliability of the detection result. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of a UAV frequency hopping signal detection method based on a clustering algorithm according to the present invention. Figure 2 This is a schematic diagram illustrating the principle of the frequency hopping signal clustering spectrum diagram of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] It is understood that the terms "first," "second," etc., used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this application, a first script may be referred to as a second script, and similarly, a second script may be referred to as a first script.

[0018] like Figure 1 As shown, a method for detecting UAV frequency hopping signals based on clustering algorithms includes: Electromagnetic signals of a preset duration are acquired, and short-time Fourier transform is performed on the electromagnetic signals to obtain spectral data. Noise filtering is then applied to the spectral data to obtain frequency hopping signal points.

[0019] In this embodiment, the electromagnetic signal acquisition system consists of a broadband antenna, an RF front-end, and a high-speed data acquisition card. The center frequency for signal acquisition is set to the 2.45GHz band commonly used in commercial drones, and the sampling bandwidth is set to 50MHz to fully cover frequency hopping signals that may occur within this band. The preset duration is, for example, 1000 milliseconds. After acquiring the electromagnetic environment signal for the preset duration, it is converted to a digital intermediate frequency (IF), and then converted from analog to digital to obtain the I / Q data stream. Short-time Fourier transform (SFT) is then used to perform time-frequency analysis on the I / Q data stream. The SFT analysis method is existing technology and will not be described here.

[0020] The obtained spectral data is specifically a spectrum graph, where the horizontal axis represents time and the vertical axis represents frequency. Since the acquired electromagnetic signals include not only the frequency-hopping signals of the target UAV but also a significant amount of background noise, noise filtering processing of the spectral data is necessary. In this embodiment, a constant false alarm rate (CFAR) algorithm is used to determine the segmentation threshold. This method is also existing technology and will not be described further here. Points with energy values ​​higher than their corresponding adaptive thresholds are marked as frequency-hopping signal points, and points lower than the thresholds are marked as noise points.

[0021] Clustering of frequency-hopping signal points generates historical observation clusters. For historical observation clusters of the same type, two sets of parent hypotheses are created. Each set of parent hypotheses includes the number of UAVs, the hypothesis confidence, and the hypothesis trajectory. The hypothesis trajectory includes the sequence formed by the historical observation clusters corresponding to each UAV in the parent hypotheses.

[0022] After the above processing steps, the frequency-hopping signal points will still contain some noise points. The UAV signal has a dwell time at each frequency-hopping point, which appears as a rectangular region with a certain time width and bandwidth on the spectrogram. To identify and locate these rectangular regions, this embodiment uses the DBSCAN algorithm for clustering. Each historical observation cluster formed after clustering represents a frequency-hopping signal point of the UAV.

[0023] It should be noted that the technical solution presented in this embodiment is based on existing technology, specifically the method proposed in the background section. After acquiring the frequency-hopping signal points, the centroids of the frequency-hopping signals are initially sorted based on the similarity of the physical characteristics of the signals and the principle of temporal exclusivity. Signals with highly similar bandwidth, waveform, and energy characteristics are initially grouped into one category, and temporal overlap is detected among signals of the same category. If overlap exists, they are identified as multiple targets, thereby obtaining a baseline estimate of the number of UAVs. Then, the UAV signals belonging to the same cluster after clustering are further analyzed using this technical solution to determine how many UAVs generated these signals. The analyzed number is then superimposed with the previously initially sorted number to obtain the final result.

[0024] This embodiment creates two sets of parent hypotheses for the same type of historical observation cluster, each set corresponding to an estimate of the number of drones. For example, parent hypothesis 1 corresponds to one drone, and parent hypothesis 2 corresponds to two drones. Then, an initial hypothesis confidence level is calculated for each set of parent hypotheses. The higher the hypothesis confidence level, the more reasonable the parent hypothesis is. The hypothesis confidence level will be continuously adjusted as the monitoring process progresses.

[0025] Assuming the trajectory represents the affiliation of historical observation clusters, for example, there are currently historical observation clusters A, B, C, and D of the same type. The current parent hypothesis is that there are two drones, 1 and 2, where the assumed trajectory of drone 1 is historical observation clusters A and C, and the assumed trajectory of drone 2 is historical observation clusters B and D. Figure 2 As shown, that is to say, Figure 2 Frequency hopping signals A and C are generated by UAV 1, while frequency hopping signals B and D are generated by UAV 2.

[0026] The current observation cluster is generated based on the re-acquired electromagnetic signals, and the current observation cluster is assigned to the parent hypothesis in different ways to obtain multiple sets of sub-hypotheses.

[0027] As time progresses, the system continuously acquires new electromagnetic signals. These new signals need to be correlated with existing drone trajectories or identified as newly emerging drones. Specifically, the current observation cluster refers to the new observation cluster obtained in the latest processing cycle by analyzing and processing newly acquired electromagnetic signals through short-time Fourier transform and clustering. The current observation cluster is the frequency-hopping signal of subsequently emerging drones. By assigning the current observation cluster to the parent hypothesis in different ways, the sub-hypotheses inherit the historical trajectory of the parent hypothesis and assign an affiliation to the current observation cluster. For example, given a parent hypothesis with two drones, after receiving a new observation cluster C, multiple sub-hypotheses can be derived, such as C belongs to drone 1, C belongs to drone 2, or C belongs to a newly emerging drone 3. Unless otherwise specified, historical observation clusters and current observation clusters are assumed to be of the same type.

[0028] Calculate the goodness of fit between the current observation cluster and the historical observation cluster for each sub-hypothesis, update the hypothesis confidence of the sub-hypothesis based on the goodness of fit, filter the sub-hypotheses based on the updated hypothesis confidence, and keep the sub-hypotheses for the next processing cycle.

[0029] Since the above process generates a large number of sub-hypotheses, most of which are unreasonable—for example, forcibly assigning a signal point that clearly does not conform to the frequency hopping pattern of a certain UAV—it is necessary to filter these sub-hypotheses. Specifically, the confidence level of each sub-hypothesis is updated by calculating the goodness of fit between the new observation cluster and the hypothetical trajectory. The goodness of fit is an indicator that measures whether the trajectory still maintains its inherent regularity after a hypothetical trajectory is added to the current observation cluster; a higher value indicates a better fit. Then, based on the original confidence level of its parent hypothesis and the current goodness of fit, the credibility of the sub-hypothesis, i.e., the hypothesis confidence level, is calculated. The specific calculation method will be introduced later.

[0030] Finally, filter out sub-hypotheses with low confidence levels. For example, set a threshold to remove sub-hypotheses with confidence levels below that threshold, or sort the sub-hypotheses according to their confidence levels and remove those that are ranked lower.

[0031] If the confidence level of a sub-hypothesis is higher than that of all other sub-hypotheses for multiple consecutive processing cycles, then that sub-hypothesis is taken as the final detection result of the frequency hopping signal.

[0032] In the initial stage of signal detection, due to limited observation data, the confidence levels of multiple hypotheses may alternate, making it impossible to determine which hypothesis is correct. Therefore, this step establishes a decision criterion. Specifically, a processing cycle value is defined, for example, 5. If a sub-hypothesis has a confidence level higher than other hypotheses for 5 consecutive processing cycles, then this sub-hypothesis is taken as the final detection result. This sub-hypothesis includes the number of drones that may exist in the current environment and their respective frequency hopping trajectory sequences.

[0033] Specifically, the system presets a maximum decision period, such as 20 processing cycles. If no sub-hypothesis is selected as the detection result of the frequency hopping signal within the maximum decision period, then after the 20th processing cycle, the top P sub-hypotheses with higher confidence at this time are output as the result, to avoid the system entering an infinite loop due to the inability to select the best sub-hypothesis, where P is an integer greater than or equal to 1.

[0034] In this embodiment, clustering frequency hopping signal points to generate historical observation clusters includes the following steps: Time-frequency distance statistics are performed on the frequency hopping signal points and a distance distribution histogram is generated. A candidate neighborhood radius set is generated based on the peak position of the distance distribution histogram. The candidate neighborhood radius set includes multiple neighborhood radius parameters. DBSCAN clustering is performed based on the candidate neighborhood radius set to obtain multiple clustering results.

[0035] Although the DBSCAN algorithm does not require a preset number of clusters, its core parameter, the neighborhood radius, has a significant impact on the clustering results. In electromagnetic environments, the signal density of different drones at different times can vary greatly. Using a fixed neighborhood radius cannot adapt to all situations simultaneously, easily leading to dense frequency-hopping signal points merging into clusters while sparse points are ignored as noise. Furthermore, real frequency-hopping signals form rectangular regions on the spectrum, typically dense at the center and sparse at the edges. To address these issues, this embodiment proposes the following technical solution: First, the k-distance distribution is calculated. Specifically, for each frequency-hopping signal point in the dataset, its Euclidean distance to its k-th nearest neighbor is calculated. This distance reflects the local point density in the region where the frequency-hopping signal point is located; the smaller the distance value, the higher the density around that point. By statistically analyzing the k-nearest neighbor distances of all points and plotting them as a histogram, the peak values ​​on the histogram correspond to the internal distances of one or more signal clusters with similar densities in the dataset. This step provides density scale information about the distribution of frequency-hopping signal points.

[0036] Next, a peak detection algorithm, such as the FindPeaks algorithm, is used to find all peak locations. The distance values ​​corresponding to all detected peaks are collected and arranged in ascending order to generate a candidate neighborhood radius in ascending order. Then, DBSCAN clustering is performed in parallel on the frequency hopping signal points using the previously generated candidate radii. Assuming there are K candidate radius parameters, k clustering result sets will be obtained.

[0037] A consensus evidence matrix is ​​constructed based on the number of times any two frequency-hopping signal points are classified into the same cluster in all clustering results. The frequency-hopping signal points are used as nodes to construct a graph structure. The edge weights between nodes in the graph structure are determined based on the consensus evidence matrix. After removing edges with edge weights less than a preset voting threshold, a corrected graph is obtained. Connectivity analysis is performed on the corrected graph, and nodes with connectivity are used as initial observation clusters. Edge correction is performed on the initial observation clusters to obtain historical observation clusters.

[0038] Create an N x N consensus evidence matrix, where N is the total number of frequency hopping signal points. Each element in the consensus evidence matrix is ​​initially set to 0. In the k clustering results, if frequency hopping signal points a and b are grouped into the same cluster in the i-th clustering result, then increment the value of the element at the a-th row and b-th column of matrix M by 1. Therefore, the higher the final score of an element in the consensus evidence matrix, the more reliable the clustering result for the corresponding frequency hopping signal point.

[0039] Next, each frequency-hopping signal point is used as a node to construct a graph structure, with the edge weight between two nodes being the corresponding element value in the consensus evidence matrix. A voting threshold is set, for example, k / 2. Edges with weights less than the voting threshold are removed from the graph structure to obtain a sparse graph. Analysis is performed on the sparse graph to determine nodes with connectivity, and the frequency-hopping signal points corresponding to nodes with connectivity are considered as a historical observation cluster.

[0040] In this embodiment, correcting the initial observation cluster includes the following steps: Unclustered frequency hopping signal points are used as target points. A fusion neighborhood is generated based on the target points. Clustered frequency hopping signal points in the fusion neighborhood are used as belonging points. If the proportion of belonging points belonging to the same initial observation cluster exceeds the belonging threshold, the target points are merged into the initial observation cluster. This step is repeated until each target point is traversed to complete the edge correction of the initial observation cluster.

[0041] In the spectrum diagram of a UAV, the density of frequency-hopping signal points at the edge of its observation clusters gradually changes. The standard DBSCAN algorithm often incorrectly labels the real frequency-hopping signal points at the cluster edge as noise, resulting in incomplete observation clusters and affecting the accuracy of subsequent tracking. Therefore, the above steps leave behind edge frequency-hopping signal points with extremely low density and random noise points, resulting in incomplete cluster boundaries. To address this, this embodiment checks the remaining points after clustering. Specifically, it traverses each un-clustered frequency-hopping signal point and sets a circular fusion neighborhood for it. The radius of the fusion neighborhood can be set according to actual needs, for example, 1.5 times the average radius of the candidate neighborhoods in the candidate neighborhood radius set. If multiple frequency-hopping signal points exist within the fusion neighborhood, the number of these frequency-hopping signal points belonging to each of the formed clusters is counted. If most of the frequency-hopping signal points belong to a formed cluster, and the ratio of the total number of frequency-hopping signal points to the total number of all frequency-hopping signal points within the radius of the fusion neighborhood exceeds a preset attribution threshold, then the noise point is recycled and merged into the cluster.

[0042] For example, suppose a blue and red cluster have already formed, and there is a black frequency-hopping signal point to be determined. The set assignment threshold is 70%. The fusion neighborhood of the black frequency-hopping signal point has 10 already assigned frequency-hopping signal points, of which 8 belong to the blue cluster, a ratio of 8 / (8+2) = 80%. Because 80% exceeds the set assignment threshold of 70%, the black noise point is determined to be an edge point of the blue cluster and is merged into the blue cluster. Through the above steps, the final generated historical observation cluster is ensured to be more complete in morphology, and the calculation of its centroid, size, and other features is more accurate.

[0043] In this embodiment, creating two sets of parent hypotheses based on historical observation clusters includes the following steps: The center frequency and time of the centroid points are extracted from the historical observation clusters in chronological order to form a frequency time series. The frequency time series is reconstructed in phase space to obtain state vectors. The state vectors are clustered based on the distance between them to obtain UAV clusters. The number of UAV clusters is determined as the number of UAVs K. The separation degree between UAV clusters is calculated. The average of all separation degrees is mapped to the first confidence degree corresponding to the number of UAVs K.

[0044] Each historical observation cluster contains time and frequency data, specifically the occurrence time and frequency of the centroid frequency hopping signal points within the cluster. The historical observation clusters are then sorted chronologically to obtain a mixed time series. In multi-UAV scenarios, adjacent signal points may belong to different UAVs, and their time intervals may not exhibit a periodic pattern. Therefore, this embodiment first calculates the time interval between any two historical observation clusters and uses binning or kernel density estimation to statistically analyze all time intervals, thereby determining the frequency distribution of the time intervals. Then, the top L most frequent time intervals are selected as candidate frequency hopping periods, reflecting potential frequency hopping patterns within the scenario.

[0045] To separate the frequency hopping pattern of each UAV from the mixed time series, this invention employs an improved time-delay embedding method for phase space reconstruction. The time-delay embedding method is a way to recover a low-dimensional observation sequence to a high-dimensional state space. Traditional embedding methods typically use a fixed time delay to select adjacent components, which is effective when processing single signals. However, when processing mixed signals from multiple UAVs, this fixed delay cannot match the actual frequency hopping period of any single UAV, leading to reconstruction failure.

[0046] Therefore, in this embodiment, after determining the optimal embedding dimension using the spurious nearest neighbor method, a multi-dimensional state vector is constructed for each sequence point in the mixed time series, starting from the end point. The construction process of the state vector is as follows: the first component of the state vector is the frequency corresponding to the end sequence point. Next, for all sequence points before the end sequence point, the time difference between the end sequence point and the end sequence point and the difference between all candidate periods are calculated, and the sequence point that minimizes this difference is selected as the predecessor point. Then, the actual time interval and actual frequency interval between the predecessor point and the end sequence point are used as the second and third components of the end sequence point. The search continues to find the predecessor point of the predecessor point, and so on, until the state vector with the optimal embedding dimension is constructed.

[0047] For example, candidate frequency hopping periods include 10ms and 17ms. The current mixed time series includes sequence points A, B, and C. Now, we need to generate a state vector for sequence point C. The time difference between sequence points B and C is 10.1ms, and the differences with the two candidate frequency hopping periods are 0.1ms and 6.9ms, respectively. The time difference between sequence points A and C is 7.2ms, and the differences with the two candidate frequency hopping periods are 2.8ms and 9.8ms, respectively. Since the 0.1ms difference is the smallest, sequence point B is determined as the predecessor of sequence point C, and sequence point B is used to continue constructing the state vector of sequence point C. If the optimal embedding dimension is 6, then 6 components need to be found for the state vector, which means finding two predecessor points before the final sequence point.

[0048] By traversing all sequence points of the mixed time series, a state vector with its own pattern as the core is generated for each sequence point, and finally a series of high-dimensional state vectors are obtained. This series of state vectors together constitutes the reconstructed phase space.

[0049] Specifically, if a sequence point fails to find a predecessor point during the iteration process, the construction of a state vector for that sequence point will be abandoned, so that the sequence point will not have a corresponding state vector in the final reconstructed phase space. This step ensures that the state vectors of subsequent comparisons all have the same dimension.

[0050] While the frequency-hopping behavior of each UAV appears complex and varied on the surface, it is essentially driven by a deterministic rule. This rule defines the trajectory of the UAV's next frequency-hopping signal. Once the reconstructed phase space is obtained using the method described above, all state vectors belonging to the same UAV, because they were dynamically matched with their common frequency-hopping period during construction and follow the same hopping rule, will inevitably converge and be constrained to a specific geometric shape in the multidimensional reconstructed phase space, which is called an attractor. The characteristic of an attractor is the iterative regression of its trajectory; that is, the frequency-hopping signal state of a UAV at time i will inevitably evolve into a frequency-hopping signal state very close to that at some future time j. If there are K UAVs, then there will necessarily be K attractors simultaneously in the reconstructed phase space. This is because, with different frequency-hopping patterns among the UAVs, the state vectors of different UAVs were matched with different candidate frequency-hopping periods during construction, causing them to occupy different regions in the phase space.

[0051] This embodiment uses a clustering algorithm to determine the number of attractors, i.e., the number of drones. All state vectors are grouped based on the distance between them. Since state vectors belonging to the same attractor are spatially close, the clustering algorithm can effectively divide all state vectors into multiple drone clusters, each corresponding to one attractor, i.e., one drone. Here, we assume the number of drones is K. The first confidence level is obtained by calculating the separation degree. Specifically, the inter-cluster distance of drone clusters can be used as the separation degree, which can be calculated using the maximum distance method, minimum distance method, or average distance method. After obtaining the separation degree between each pair of drone clusters, their average value is taken and mapped to the range of 0-1 to obtain the separation degree representing the overall separation. The higher the separation degree, the more distinct the attractors are in phase space, and the more reliable the hypothesis K is.

[0052] Calculate the state transition probability between drone clusters, determine the transition path based on the state transition probability, correct the number of drones K based on the number of clusters included in the transition path, obtain the number of drones I, and calculate the second confidence level corresponding to the number of drones I based on the average state transition probability in the transition path.

[0053] Generate hypothetical trajectories corresponding to the number of drones K and the number of drones I. Use the number of drones K, the first confidence level, and the corresponding hypothetical trajectory as the first set of parent hypotheses, and use the number of drones I, the second confidence level, and the corresponding hypothetical trajectory as the second set of parent hypotheses.

[0054] In reality, for complex and variable frequency-hopping signals, one possibility is that they are generated by multiple drones, and another possibility is that they are generated by only a few drones through continuous switching of hopping modes, that is, continuously switching the hopping time or frequency. The first parent hypothesis generated by the above technical solution considers the first possibility. Subsequently, a second parent hypothesis needs to be generated to consider the second possibility. The following example illustrates the process of generating the second parent hypothesis. First, each state vector is labeled with its corresponding drone cluster number, for example, drone clusters A, B, C... Then, attractor transition sequences are constructed. If the mixed time series s1, s2, s3, s4, ... belong to attractors A, D, B, C, A, ... respectively, then the generated drone cluster transition sequence is A, D, B, C, A, ...

[0055] Next, the state transition probabilities between different UAV clusters in the transition sequence are calculated. The calculation method for the state transition probabilities is existing technology and will not be described here. Assuming P(A, B) = 0.9, P(B, C) = 0.8, and P(C, A) = 1, if the first critical value is 0.7, then the above calculation results indicate that there is a high probability of a transition path between UAV clusters A, B, and C. That is, after UAV cluster A appears, it is highly likely that UAV cluster B will appear; after UAV cluster B appears, it is highly likely that UAV cluster C will appear; and after UAV cluster C appears, it is highly likely that UAV cluster A will appear. This confirms that a single UAV is cyclically switching between three frequency hopping modes. Therefore, UAV clusters A, B, and C are identified as one transition path. If only one such transition path exists, then the three UAV clusters are merged into one, and the remaining number of UAV clusters is I = K - 3 + 1.

[0056] In this embodiment, calculating the second confidence level corresponding to the number of drones I based on the average state transition probability in the transition path includes the following steps: The UAV clusters included in the transition path are taken as target clusters. The state transition probabilities between target clusters are organized into a numerical set. The average value of the numerical set is calculated to obtain the second confidence level.

[0057] When calculating the second confidence level, if multiple independent transition paths exist, the average state transition probability of all transition paths is calculated, and the first confidence level is multiplied by this average to obtain the second confidence level. For example, if it is determined that there are two transition paths, one of which requires merging UAV clusters A, B, and C into one, and the other requires merging UAV clusters D and E into one, then the target clusters are A, B, C, D, and E, where P(A, B) = 0.9, P(B, C) = 0.8, P(C, A) = 1, P(D, E) = 0.7, and P(E, D) = 0.6. These five values ​​constitute the set of state transition probabilities between the target clusters, and the average value of the set of values ​​is (0.9 + 0.8 + 1 + 0.75 + 0.8) / 5 = 0.85.

[0058] In this embodiment, generating the hypothetical trajectories corresponding to the number of drones K and the number of drones I includes the following steps: Obtain the historical observation clusters included in each UAV cluster in the parent hypothesis, arrange the historical observation clusters in chronological order to obtain the trajectory groups of the corresponding UAV clusters, allocate and integrate the trajectory groups to obtain the hypothetical trajectories corresponding to the number of UAVs K and the number of UAVs I.

[0059] For example, assuming a parent hypothesis of 2 drones (K), drone clusters 1 and 2 are obtained through phase space reconstruction clustering. First, the historical observation clusters C1, C3, and C5 contained in drone cluster 1 are obtained and sorted chronologically to obtain trajectory group 1: C1, C3, C5. Similarly, the historical observation clusters {C2, C4} contained in drone cluster 2 are obtained and sorted to obtain trajectory group 2: C2, C4. Finally, the sets [C1, C3, C5] and [C2, C4] of trajectory group 1 and trajectory group 2 are used as the complete hypothetical trajectories when K=2 in the parent hypothesis, representing the historical frequency hopping sequences of the two drones.

[0060] In this embodiment, calculating the goodness of fit between the current observation cluster and the historical observation cluster in each sub-hypothesis, and updating the hypothesis confidence of the sub-hypothesis based on the goodness of fit includes the following steps: Calculate the time interval sequence and frequency interval sequence of historical observation clusters in the parent hypothesis. Calculate the actual time interval and actual frequency interval between the current observation cluster and each historical observation cluster in each hypothesis trajectory. Count the number of occurrences of each actual time interval and actual frequency interval in the time interval sequence and frequency interval sequence. Select the highest occurrence count as the time interval occurrence count and frequency interval occurrence count. Calculate the goodness of fit based on the time interval occurrence count and frequency interval occurrence count. Combine the goodness of fit with the parent hypothesis using weighted fusion to obtain the updated confidence score. Normalize the updated confidence scores of all sub-hypotheses to obtain the hypothesis confidence scores of the sub-hypotheses.

[0061] First, obtain the historical observation clusters arranged chronologically from the parent hypothesis. Calculate the time difference and frequency difference between any two adjacent historical observation clusters to construct a historical time interval sequence and a historical frequency interval sequence. Count the total number of elements in each sequence as the length of the time series and the length of the frequency series, respectively. Next, obtain the currently acquired observation cluster and calculate the actual time interval and actual frequency interval between the current observation cluster and each historical observation cluster in each sequence of the parent hypothesis. For example, assuming the trajectory includes historical observation clusters C1, C3, and C5, and the current observation cluster is C10, calculate the time interval between C1 and C10, the time interval between C3 and C10, and the time interval between C5 and C10. The frequency interval is calculated similarly.

[0062] The occurrence frequency of each actual time interval in the time interval sequence is counted, and the highest occurrence frequency is selected. For example, if an actual time interval occurs 10 times in the time interval sequence, and the occurrence frequency of other actual time intervals is less than 10 times, then 10 times is the highest occurrence frequency of the time interval. Similarly, the highest occurrence frequency of the frequency interval is obtained. The highest occurrence frequency of the time interval is divided by the length of the time series to obtain the time matching rate, and the highest occurrence frequency interval is divided by the length of the frequency series to obtain the frequency matching rate. The average of the time matching rate and the frequency matching rate is determined as the goodness of fit between the current observation cluster and the historical observation clusters. In particular, if the highest occurrence frequency of both time and frequency is 0 after calculation, it means that the current observation cluster does not match the historical observation clusters. In this case, the subsequent matching rate is not calculated, and the current observation cluster is treated as a new target for monitoring.

[0063] Set a first weight and a second weight. Multiply the original confidence of the parent hypothesis by the first weight and multiply the good fit by the second weight. Add the two products to obtain the updated confidence of the sub-hypothesis. After calculating the updated confidence of all sub-hypotheses, normalize the updated confidence of all sub-hypotheses to obtain the hypothesis confidence of the sub-hypothesis.

[0064] This invention also provides a UAV frequency hopping signal detection system based on a clustering algorithm. This system implements the above-described method and includes: The filtering module acquires electromagnetic signals for a preset duration, performs a short-time Fourier transform on the electromagnetic signals to obtain spectral data, and performs noise filtering on the spectral data to obtain frequency hopping signal points.

[0065] The first analysis module clusters the frequency hopping signal points to generate historical observation clusters. For the same type of historical observation cluster, two sets of parent hypotheses are created. Each set of parent hypotheses includes the number of UAVs, the hypothesis confidence level, and the hypothesis trajectory. The hypothesis trajectory includes the sequence formed by the historical observation cluster corresponding to each UAV in the parent hypothesis.

[0066] The second analysis module generates the current observation cluster based on the re-acquired electromagnetic signals, assigns the current observation cluster to the parent hypothesis in different ways to obtain multiple sets of sub-hypotheses, calculates the fit degree between the current observation cluster and the historical observation cluster in each sub-hypothesis, updates the hypothesis confidence of the sub-hypothesis based on the fit degree, filters the sub-hypotheses based on the updated hypothesis confidence, and retains the sub-hypotheses to enter the next processing cycle.

[0067] The output module takes a sub-hypothesis as the final detection result of the frequency hopping signal if the hypothesis confidence of the sub-hypothesis is higher than that of all other sub-hypotheses for multiple consecutive processing cycles.

[0068] It should be further noted that those skilled in the art will understand that the preset parameters mentioned in the above specific embodiments, such as the specific values ​​of "preset voting threshold," "attribution threshold," "first weight and second weight," "multiple consecutive processing cycles," and "maximum decision cycle," are not strict limitations of the present invention. The specific values ​​of these parameters can be flexibly set and dynamically optimized according to the complexity of the electromagnetic environment in the actual application scenario, the signal characteristics of the target UAV, and long-term empirical data to achieve the best detection effect.

[0069] Furthermore, the core objective of the method and system provided by this invention is to accurately estimate and determine the number of drone targets in complex and easily confused scenarios. Therefore, the final detection results output should be considered as an important technical reference and decision-making basis, aiming to provide key data support for subsequent accurate drone identification, threat level assessment, and the formulation of effective control or interference countermeasures, rather than an absolute guarantee of the number of targets in the physical world.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

[0072] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting frequency hopping signals of unmanned aerial vehicles (UAVs) based on clustering algorithms, characterized in that, include: Collect electromagnetic signals for a preset duration, perform short-time Fourier transform on the electromagnetic signals to obtain spectral data, and filter noise from the spectral data to obtain frequency hopping signal points; The frequency hopping signal points are clustered to generate historical observation clusters. For the same type of historical observation cluster, two sets of parent hypotheses are created. Each set of parent hypotheses includes the number of UAVs, the hypothesis confidence, and the hypothesis trajectory. The hypothesis trajectory includes the sequence formed by the historical observation clusters corresponding to each UAV in the parent hypotheses. The current observation cluster is generated based on the electromagnetic signals acquired again, and the current observation cluster is assigned to the parent hypothesis in different ways to obtain multiple sets of sub-hypotheses; Calculate the goodness of fit between the current observation cluster and the historical observation cluster for each sub-hypothesis, update the hypothesis confidence of the sub-hypothesis based on the goodness of fit, filter the sub-hypotheses based on the updated hypothesis confidence, and keep the sub-hypotheses for the next processing cycle. If the confidence level of a sub-hypothesis is higher than that of all other sub-hypotheses for multiple consecutive processing cycles, then that sub-hypothesis is taken as the final detection result of the frequency hopping signal.

2. The method according to claim 1, characterized in that, Clustering the frequency hopping signal points to generate historical observation clusters includes the following steps: Time-frequency distance statistics are performed on the frequency hopping signal points and a distance distribution histogram is generated. A candidate neighborhood radius set is generated based on the peak position of the distance distribution histogram. The candidate neighborhood radius set includes multiple neighborhood radius parameters. DBSCAN clustering is performed based on the candidate neighborhood radius set to obtain multiple clustering results. A consensus evidence matrix is ​​constructed based on the number of times any two frequency-hopping signal points are classified into the same cluster in all clustering results. The frequency-hopping signal points are used as nodes to construct a graph structure. The edge weights between nodes in the graph structure are determined based on the consensus evidence matrix. After removing edges with edge weights less than a preset voting threshold, a corrected graph is obtained. Connectivity analysis is performed on the corrected graph, and nodes with connectivity are used as initial observation clusters. Edge correction is performed on the initial observation clusters to obtain historical observation clusters.

3. The method according to claim 2, characterized in that, Edge correction for the initial observation cluster includes the following steps: The frequency hopping signal points that have not been clustered are taken as target points. A fusion neighborhood is generated based on the target points. The frequency hopping signal points that have been clustered in the fusion neighborhood are taken as belonging points. If the proportion of belonging points belonging to the same initial observation cluster exceeds the belonging threshold, the target points are merged into the initial observation cluster. This step is repeated until each target point is traversed to complete the edge correction of the initial observation cluster.

4. The method according to claim 1, characterized in that, Creating two sets of parent hypotheses for the same type of historical observation clusters involves the following steps: The frequency and time of centroid points are extracted from historical observation clusters in chronological order to form a frequency time series. The frequency time series is reconstructed in phase space to obtain the state vector of each historical observation cluster. The historical observation clusters are clustered based on the distance between the state vectors to obtain UAV clusters. The number of UAV clusters is determined as the number of UAVs K. The separation degree between UAV clusters is calculated. The average value of all separation degrees is mapped to the first confidence degree corresponding to the number of UAVs K. Calculate the state transition probability between drone clusters, determine the transition path based on the state transition probability, correct the number of drones K based on the number of clusters included in the transition path, obtain the number of drones I, and calculate the second confidence level corresponding to the number of drones I based on the average state transition probability in the transition path. Generate hypothetical trajectories corresponding to the number of drones K and the number of drones I. Use the number of drones K, the first confidence level, and the corresponding hypothetical trajectory as the first set of parent hypotheses, and use the number of drones I, the second confidence level, and the corresponding hypothetical trajectory as the second set of parent hypotheses.

5. The method according to claim 4, characterized in that, Calculating the second confidence level corresponding to the number of drones I based on the average state transition probability in the transition path includes the following steps: The UAV clusters included in the transition path are taken as target clusters. The state transition probabilities between target clusters are organized into a numerical set. The average value of the numerical set is calculated to obtain the second confidence level.

6. The method according to claim 4, characterized in that, Generating the hypothetical trajectories corresponding to the number of drones K and I includes the following steps: Obtain the historical observation clusters included in each UAV cluster in the parent hypothesis, arrange the historical observation clusters in chronological order to obtain the trajectory groups of the corresponding UAV clusters, allocate and integrate the trajectory groups to obtain the hypothetical trajectories corresponding to the number of UAVs K and the number of UAVs I.

7. The method according to claim 1, characterized in that, Calculating the goodness of fit between the current observation cluster and the historical observation cluster for each sub-hypothesis, and updating the hypothesis confidence of the sub-hypothesis based on the goodness of fit, includes the following steps: Calculate the time interval sequence and frequency interval sequence of historical observation clusters in the parent hypothesis. Calculate the actual time interval and actual frequency interval between the current observation cluster and each historical observation cluster in each hypothesis trajectory. Count the number of occurrences of each actual time interval and actual frequency interval in the time interval sequence and frequency interval sequence. Select the highest occurrence count as the time interval occurrence count and frequency interval occurrence count. Calculate the goodness of fit based on the time interval occurrence count and frequency interval occurrence count. Combine the goodness of fit with the parent hypothesis using weighted fusion to obtain the updated confidence score. Normalize the updated confidence scores of all sub-hypotheses to obtain the hypothesis confidence scores of the sub-hypotheses.

8. A UAV frequency hopping signal detection system based on a clustering algorithm, used to implement the method as described in any one of claims 1-7, characterized in that, The system includes: The filtering module acquires electromagnetic signals for a preset duration, performs a short-time Fourier transform on the electromagnetic signals to obtain spectral data, and performs noise filtering on the spectral data to obtain frequency hopping signal points. The first analysis module clusters the frequency hopping signal points to generate historical observation clusters. For the same type of historical observation cluster, two sets of parent hypotheses are created. Each set of parent hypotheses includes the number of UAVs, the hypothesis confidence level, and the hypothesis trajectory. The hypothesis trajectory includes the sequence formed by the historical observation clusters corresponding to each UAV in the parent hypothesis. The second analysis module generates the current observation cluster based on the re-acquired electromagnetic signals, assigns the current observation cluster to the parent hypothesis in different ways to obtain multiple sets of sub-hypotheses, calculates the fitting degree between the current observation cluster and the historical observation cluster in each sub-hypothesis, updates the hypothesis confidence of the sub-hypothesis based on the fitting degree, filters the sub-hypotheses based on the updated hypothesis confidence, and retains the sub-hypotheses to enter the next processing cycle. The output module takes a sub-hypothesis as the final detection result of the frequency hopping signal if the hypothesis confidence of the sub-hypothesis is higher than that of all other sub-hypotheses for multiple consecutive processing cycles.

Images