A New Frequency Hopping Signal Detection of Civil UAV Based on Improved K-Means Clustering Algorithm

The rapid development of civil UAV promotes the social and economic development, and the frequent “flying illegally” events has brought great challenges to aviation safety and government supervision. The frequency hopping communication system used in UAV data transmission and control link has the advantages of anti-jamming and anti-interception, and its complex electromagnetic environment, which also brings great difficulties to UAV detection. In this paper, the detection of civil UAV is realized by frequency hopping signal monitoring. Firstly, by analyzing the signal characteristics of UAVs, an adaptive noise threshold calculation method is proposed for find the signals from spectrum data. Then, the improved clustering analysis algorithm is proposed based on constructed the waveform shape characteristics and peak characteristics of UAV frequency hopping signal. Finally, according to the designed experimental process, the experimental environment is set up, and the UAV monitoring, discovery and parameter estimation are realized by using the improved clustering analysis algorithm, and compared with K-means, K-means++, DBSCAN, Multi-hop and Auto-correlation methods. The results show that the method has certain robustness and has a good application prospect.


I. INTRODUCTION
Unmanned aerial vehicles (UAVs) of civil applications is growing rapidly across many application domains and improving ones' quality of life, such as real-time monitoring, providing wireless coverage, remote sensing, search and rescue, delivery of goods, security and surveillance, precision agriculture, and civil infrastructure inspection, up till now, UAVs industry has greatly promoted social and economic development, and become thousands of billions market value [1]- [3]. Meanwhile, civil UAVs are also faced many challenges within specific vertical domains and across many application domains, such as UAVs amateurs have weak air safety awareness, immature operation and imperfect air control system, and ''illegal flying'' of civil UAVs have brought unprecedented threats to aviation safety and personal The associate editor coordinating the review of this manuscript and approving it for publication was Jing Liang . privacy [4]. To overcome these challenges, formulating relevant laws, regulations and policies to implement the supervision of UAVs must be done by the government, as well as technicalities of detecting, locating and controlling UAVs are urgently developed [5].
In existing studies, UAV detection methods mainly include sound monitoring, visible light, infrared or radar detection and radio spectrum monitoring, etc. [3]. Among them, sound monitoring is susceptible to the noise of surrounding life, visible light is usually affected by weather and building occlusion. While UAV miniaturization and even ''invisible'', makes infrared and radar detection technology difficult to play an advantage. Radio spectrum monitoring is through equipment (radio receiver and antenna, etc.) to monitor the UAV frequency band all-weather, analyze the spectrum data, then detect the control signal or the communication signal. Compared with these methods, spectrum monitoring has the advantages of good applicability and less influence by the environment, become an important method for UAV discovery [6]- [10], [28].
The electromagnetic environment of UAV in radio frequency band is complex, noise and interference coexist [11], and most of the UAV signals are frequency hopping (FH) signals, which further increases the difficulty of detection and analysis. Research on FH signal detection, separation and recognition based on spectrum monitoring data has become a hot spot in UAV signal analysis field [12]- [15]. FH signal detection methods are mainly frequency domain and time domain detection. Common frequency domain analysis methods include Fourier transform [16], wavelet analysis [17], [18], Wigner-Ville distribution [19], spectrogram, etc. [20]- [22]. Time domain detection methods include image analysis, energy detection, correlation detection, etc. [23]- [28].
The paper in [14] propose an improved channelized MWC scheme for non-cooperative detection and frequency estimation of FHSS signals, by adjusting the channel weighting factor, it can better adapt to the single hop width, greatly reduce the amount of calculation, and achieve good results compared to the traditional method, but only based on theoretical analysis and numerical simulations, and not real implementation of channelized MWC. In [15], [16], [19] papers, a novel method based on sparse linear regression to solve the problem of computation burden in parameters estimate, which the results are only better on hop timing estimation, the estimation performance of other parameters is general. An approach based on frequency difference and one-dimensional non-linear filter has been proposed in [18], which the robustness against white Gaussian noise has been enhanced but need SNR is greater than 13 dB which is an extremely harsh precondition, and is also reflected in [20], the author propose a detection scheme based on the Neyman-Pearson test, and the results shows better than method of auto-correlation-based when the hopping period is short, and so on. the front papers methods almost set a harsh precondition, or manual parameter, and also not make a real environment experiment within multiple frequency hopping signals active, lacking of persuasion of practical application.
In this paper, by focusing on frequency characteristics of the civil UAV control link signals (which are FH signals), a new FH signal detection of civil UAVs based on improved K-means clustering algorithm is proposed to fast discovery civil UAVs. Concretely, an adaptive noise threshold calculation method is firstly provided to preprocess spectrum data of civil UAVs in radio frequency band. Then an improved K-means clustering algorithm is proposed to analyze FH signals of civil UAVs and estimate parameters of them. Experiment results and comparison with K-means, K-means ++ , DBSCAN, Multi-hop and Auto-correlation methods finally show that the proposed method is useful and effective to detect civil UAVs.
The rest of this paper is structured as follows: In Section II-A, characteristics of frequency hopping signals of civil UAVs are briefly reviewed. In Section II-B, introduce the idea of frequency hopping signal detection of the paper. In Section III, the process of the method to detect FH signals of civil UAVs is described in detail. In Section IV, several experiments are built to verify the performance of the proposed algorithm. In Section V, the performance of the algorithm is compared and analyzed, and conclude the paper is in Section VI to show.

II. FREQUENCY CHARACTERISTICS AND DETECTION OF CIVIL UAV CONTROL LINK SIGNALS
FH signals of civil UAVs work in open ISM band 2400MHz-2483.5MHz, in which there are WIFI, wireless mouse, cord-less phone, blue-tooth, microwave and ZigBee besides FH signals of UAVs [11], hence electromagnetic environment is often very complex, such as Figure 1 shows the spectrum waterfall map for electromagnetic environment of UAVs in ISM band at some time. In the section, frequency characteristics and detection of FH signals of civil UAVs are analyzed.

A. FREQUENCY CHARACTERISTICS OF FH SIGNALS
Comparison with other signals in the ISM band, FH signals of civil UAVs own many interesting and important characteristics, which can help us to parameterize the FH signals not only from the theoretics but also applications point of view [11], [16], [22], [28]- [32]. In the paper, as Figure 2 shows, these characteristics are summarized as follows: The bandwidth of the FH signal of a civil UAV is stable. The bandwidth of a FH signal is consisted by the transmission bandwidth and the protection bandwidth. The former is the distance between the maximum and minimum frequency points of the FH signal, which is theoretically about 2MHz; the latter is to protect the effective transmission of VOLUME 9, 2021 a FH signal. In most real applications, 1MHz protection bandwidth is added to the left and right of the FH signal respectively, hence the bandwidth of an actual received civil UAV signal is about 4MHz.
The frequency sequence of a civil UAV FH signals is fixed. The frequency sequence is a set of non-repetitive chronological frequency points of an actual received civil UAV signal. And in real applications, the FH frequency sequence of a civil UAV signal is often unchanged.
FH signals law of civil UAV can be observed. Despite the dwell time of the signals is very short between two FH in a certain span (it is usually microseconds), the FH period of the civil UAV signal is fixed, i.e., the sum of the time that the signal energy starts to disappear and the next FH energy starts to rise is fixed. In addition, FH speed (the rate of the frequency hops and time) and peak energy of a civil UAV signal are stable, the signal of 100-1000 hops per second is medium speed FH signal.
Theoretically, the aforementioned frequency characteristics can be utilized to distinguish FH signals of civil UAVs from signals in open ISM band 2400MHz-2483.5MHz. However, obtaining these frequency characteristics are not trivial. In real world applications, special data mining methods are needed to extract these frequency characteristics of FH signals of civil UAVs.
It is worth noting that there are Bluetooth signals almost have some same characteristics, such as frequency hopping, active in 2400MHz-2483MHz, frequency hopping timesequence, and a stabled dwell time, the difference is that Bluetooth signals is only 1M bandwidth in real applications, and transmission distance is less than 30 meters, and even if in an open environment, the transmission distance is less than 60 meters, which is within sight. Therefore, we can combine the actual environment and frequency hopping signal detection results for comprehensive analysis to determine whether there are Bluetooth signals.

B. DETECTING FH SIGNALS OF CIVIL UAVS
From the data mining point of view, based on frequency characteristics of a FH signal of a civil UAV, the FH signal can be parameterized, i.e., the FH signal can be expressed into frequency characteristics space, and then the civil UAV can be discovered by detecting the FH signal in the frequency characteristics space. In this paper, bandwidth, waveform, short-time peak energy and time sequence features are selected as frequency characteristics of FH signals, furthermore to extract four frequency characteristics of signals, an improved K-means clustering algorithm are proposed to classify signals in open ISM band according to each frequency characteristic, respectively. Then FH signals are detected according to clustering results of four frequency characteristics of signals. Intuitively, Figure 3 shows dwell time, bandwidth, cycle and interval of FH signals of civil UAVs.
In addition, because there exists noise in any radio signals, an adaptive noise threshold calculation method is provided  to preprocess signals. From the data analysis point of view, the adaptive noise threshold calculation method is mainly used to extract the spectrum data higher than the noise threshold, the process is equal to obtain effective spectrum data of radio signals.
From these ideas, the FH signal of UAV is detected from the characteristics of signal duration continuity, signal frequency point connectivity, bandwidth similarity, waveform feature similarity, short-time peak energy similarity and time sequence correlation. Specifically, aiming at FH signal detection in low SNR electromagnetic environment, according to the characteristics of noise and signal distribution in frequency domain, the adaptive noise threshold calculation method is used to extract the spectrum data higher than the noise threshold. Then, the improved clustering algorithm is used to cluster the extracted signals with bandwidth, waveform, energy and dwell time characteristics to detection the FH signals of UAV, and estimate the parameters. The schematic chart is shown in Figure 3, and the algorithm processing is proposed is shown in Figure 4, which shows the process of using improved clustering algorithm to detect and sorting the FH signal of UAV in detail, and analyzes the system error at the end of experiment: 1) The FH signal of UAV is extracted by adaptive threshold calculation method.
2) Using the improved clustering algorithm to extract the signal bandwidth, wave form characteristics and resident time, energy, then clustering.
3) Separation of various classification according to the time sequence, the class whose number of signals in the sorting results is less than 10% of the total number of detected signals is discarded; 4) The FH signal parameters of the classification results are estimated. 53192 VOLUME 9, 2021

III. THE PROPOSED METHOD TO DETECT FH SIGNALS OF CIVIL UAVS
According to ITU-R P372 proposal, radio noise obeys the normal distribution under certain conditions, theoretically, this also means that the number of noise points in the spectrum distribution is more than the number of signal points, and with the accumulation of time, the statistical distribution characteristics of signal and noise become more obvious. Intuitively, noise points and signal points in frames of spectrum data are convex distribution, which are shown in Figure 5, where θ is the signal threshold.

A. ADAPTIVE THRESHOLD CALCULATION METHOD
Up till now, there are many researches for how to obtain signal threshold θ [23]- [25]. In [23], an automatic algorithm extraction of the signal components (useful information) from the noisy mixture is based on the time-frequency distribution segment to a fixed number of data classes is proposed, which depends on the differences in the time frequency structures of noise and signal components. In [24] paper innovatively proposed a dynamic threshold based target signal cooperative extraction method for high frequency electromagnetic environment measurement, which reduces the influence of manual factors and improves the automatic degree and efficiency of high frequency electro-magnetic environment measurement, but it's not completely applicable to detect uav signals which is in a moving status and complex frequency environment.
In the paper, inspired by the characteristics of the FH signal of UAV, the adaptive threshold calculation method based on current spectrum data is proposed. Formally, let a frame of spectrum data in open ISM band S t = {s 1 , s 2 , · · · , s n } (where s i is signal energy level (dbµV) of i sampling point), then the signal level interval of the spectrum data can be calculated, i.e., [S min , S max ] in which S min = min{s 1 , s 2 , · · · , s n }, S max = max{s 1 , s 2 , · · · , s n }. Suppose split width of interval is b, for example, b = 2dbµv in real applications, then the level interval [S min , S max ] can be divided by m = S max −S min b sub-intervals, denoted by [S min −1 , S 1 ],[S 1 , S 2 ], · · · ,[S m-1 , S m = (S max )] such that S j = S j−1 + b for each j ∈ {2, · · · , m}. According to m subintervals, statistical analysis of the FH signal can be finished in the frame of spectrum data, then the number of sampling points in each sub-interval can be obtained, i.e., S i ∈ [S j−1 , S j ]. Accordingly, statistical law of the FH signal can be described by the following function: y is the number of sampling points in x. Based on frequency characteristics and our experiment on FH signals of UAV, the function can be shown in Figure 6, and signal threshold θ can be obtain in the function, i.e., θ is the minimum of the function. Algorithm 1 can be designed to obtain threshold θ of FH signals of UAV.

Algorithm 1 Obtaining Threshold θ of FH Signals UAV
Input: Spectrum data in open ISM band Output: Threshold θ 1) let a frame of spectrum data: break; 15) end 16) return θ Signal extraction: spectrum signals, and uav signals in special have the statistical characteristics of spectrum data as shown in the figure 5, except signals submerged in the noise, we can get a better adaptive threshold θ by Algorithm in a lower SNR environment, then the spectrum data is reprocessed based on this threshold, and signals list is obtained, then cluster analysis algorithm is carried out.

B. AN IMPROVED K-MEANS CLUSTERING ALGORITHM
Unlike the traditional K-means algorithm, in which the number of clusters need to be fixed and initial clustering centers VOLUME 9, 2021 are randomly selected, in this paper, an improved K-means clustering algorithm is proposed, in which the number of clusters are decided by density characterization of objects and clustering centers are selected as objects with maximum density.
Step 1: Given m data objects m, n is the attribute dimension of the data, the Euclidean distance of any two objects o i and o j is as follows: For object o i , denote the following distances: Intuitively, d i min is the minimum distance from all objects o j (j = i) to object o i , d 1 and d 2 are the minimum and maximum values of {d i min |i ∈ {1, 2, · · · , m}}. Based on d 1 and d 2 , the clustering radius R is calculated as follows: it means that there is at least a cluster with only one object. In real world applications, w is often in (0, 1). Obviously, the clustering radius R directly affects the result of data classification. The smaller clustering radius R is, the fewer number of elements in each cluster will be, and the greater number of clusters is. The result of rough clustering and excessive clustering is unreasonable. Therefore, we need to analyze the actual data characteristics and the distribution characteristics of the data, then select a suitable weight w and obtain the clustering results, in this paper, take w = 0.5, which get a better signal detect accuracy performance in experiment environment as shown in Figure 7. In Figure 7, w performance in experiments is trend to be stable above 80% when the value is less than 0.9, more than 0.9 decline to 75%, which may be related to the current density of spectrum data of itself, and get the best signal detect accuracy about 0.5 reach 92%.
Step 2: Calculate the density of every object in according to the clustering radius R. Formally, the density of objects o i is calculated by c 1 .
Step 3: clustering centers and Clustering results. Formally, 1) The first clustering center c 1 is selected according to the maximum density, i.e., c 1 = o l if o l is such that ρ l = max{ρ j |j = 1, 2, · · · , m} 2) Suppose k-th clustering center c k and k-th cluster C k are selected.
3) Then (k+1)-th clustering center c k+1 and (k+1)-th clus- In addition to all the element objects that have been classified, repeat the above steps for the remaining data until all the data element objects are classified.  B. EXPERIMENTAL PROCESS 1) CALCULATION OF ADAPTIVE THRESHOLD 1) Calculate the maximum level L max and minimum level L min of the spectrum data frame. The level values are segmented by 2dBµV in interval [L min , L max ], the continuous d k ∈ D, k = 1, 2, · · · m frames data are segmented in turn. The number of data points of d k in each segment is counted as {c k1 , c k2, · · · , c kp } respectively, where p is the total number of segments d k , and the segment of the maximum number of data points is obtained;

IV. EXPERIMENTAL VERIFICATION
2) Find the θ position in the interval obtained in step 1) as shown in Figure 5. If the number of data points in the interval is continuously less than the number of points in the two adjacent intervals, the upper boundary value of the interval is the signal threshold of the current frame data. As shown in Figure 9, it is a waterfall map of 1442 frames of spectrum, in which the left is the original spectrum and the right is the spectrum data of extracted signals through threshold processing and noise be cleared. The signal frequency band is 2.4GHz-2.48GHz, and the sampling time is 1s. The color changes from red RGB (255,0,0) to green RGB (0,255,0), corresponding to the maximum and minimum level value range of 1442 frames.
For FH signal detection of UAV, generally, the electromagnetic environment in a short time period will not change too much, then the initial threshold value can be used as the signal judgment for the monitoring process.

2) A SINGLE FREQUENCY SIGNAL DETECTION
1) The threshold is used to compare the level of the spectrum frame by frame. When the level is greater than the threshold, indicate signal present, otherwise no signal.
2) Continuously compare the data frame of the current frequency. When the signal exists continuously, count the start time and end time of the signal: the start time is the first time that the frequency point level is greater than the threshold, and the level first time less than the threshold as the end time of the signal after appears continuously. Denote the signal of this frequency point as s i .
3)When processing data at the last frame of the sampled data and the signal of the current frequency point still exists, the signal is discarded directly which purpose is to avoid parameter estimation error. The signal information of a single frequency expressed as S i = {S 1 , S 2 , · · · , S n } is shown in Table 1. The data filled with a blue background show that the frequency point level is greater than the threshold at the corresponding time. Each row represents a frame of data. Columns f i (MHz) indicates whether the frequency point presence signal at different times, 1 is present, and 0 indicates disappear. Frequency f i can calculated by subject: f i = 2400 + i × 0.025, of which i is serial number of f i .

3) TIME CONTINUITY DETECTION
In general, the single-hop FH signal has a certain dwell time. Intuitively expressed as at least two consecutive frames of spectrum data. For 10000 values at each frequency point, we eliminate these abnormal values whose current value is 1 and its adjacent values are 0. And the signal frame is eliminated which the end time is the last frame of sampling data. This method of time continuity detection can eliminate noise or interference to a large extent, but the low SNR signal is not eliminated. For the data in Table 1, to be eliminated is shown in the gray background, and the blue background is the reserved signal after elimination, as shown in Table 2. The waterfall map of spectrum and after eliminating noise is shown in Figure 10, with the same color mark as above.

4) MERGE OF CONNECTED REGION
For S = {S 1 , S 2 , S 3 , · · · , S i , · · · , S 3200 }, from S 1 to S k , k ∈ [1, 3200], judge whether signals S k and S k+1 are overlapping in their time intervals [t 1 , t 2 ] S k and [t 1 , t 2 ] S k+1 in turn. If so, merge the two signals in the way of maximizing the time gap between starting and ending. As shown in Table 3

5) SIGNAL BANDWIDTH DETECTION
Signal bandwidth detection steps are as follows: 1) For each signal S in the set , Bandwidth B = f e − f s , f s and f e are the starting and ending frequencies of the S , respectively.
2) If the signal bandwidth is outside the range of 0.5 MHz-10MHz, the signal is discarded, and only the signal within the range of 0.5 MHz-10MHz is retained for further feature detection.
3) The improved clustering algorithm is used for clustering and sorting, and the signal bandwidth is taken as input data. 4) Discard the classification with less than 5 elements in the result of clustering.  Figure 11 shows the results of signal bandwidth clustering and sorting. The signal bandwidths are extremely close to the same category, and the signal bandwidth has a certain gap to the different categories. The signal bandwidth detection method can effectively divide the signal data of fixed bandwidth into one category, and different bandwidth into different categories. A total of 4 categories are obtained, as shown in Figure 12, all signals of the same category are marked by the same color. From Figure 12, we can see that the first category (a) has the largest number of data elements, in fact, the signal threshold is almost close to the base of the noise, so it contains short noises. Some signals in the second category (b) can be clearly seen appear regularly, that is, there are FH signals and some interference signals. The signals in the third classification 53196 VOLUME 9, 2021 (c) have no obvious regularity, some may be burst signals, frequency hopping signals or noise data. A small number of signals in the fourth classification (d) may be signal or interference data. In subsequent clustering and parameter estimation, the classification (d) that does not have obvious FH signal characteristics and the number of elements in the category is less than 5 will be eliminated.

6) SIGNAL WAVEFORM DETECTION
The waveform characteristic value is calculated as follows: 1) ∀S ∈ , Divide the bandwidth B of S into 8 segments as 2) Calculate the mean of level value in each segment in F i one by one, get set V i = {v 1 , v 2 , · · · , v 8 }, that is, the waveform characteristic value of the signal.
3) The improved clustering algorithm is used for input parameters of all categories of signals waveform, and the results are shown in Figure 13. Specifically, for each category in Figure 13. Clustering detection result of class (a) in Figure 12 based on waveform similarity is shown in Figure 14(a1). The result of eliminating classes with elements lower than 5 is shown in Figure 14 (a2). Accordingly, the results of class (b-c) of clustering and results of eliminating the number of elements below the threshold are shown in Figure 14(b1-b2) and (c1-c2). After eliminating part of the interference, we can from the following figures intuitively find the signals in each class are further divided after some of the interference signals are eliminated. Class (c2) has no signal information after processing, and will not be processed later.

7) SIGNAL ENERGY DETECTION
Generally, during sampling, the peak energy of FH signals is relatively stable, and its energy value of is often larger in the central frequency of the signal. The further away from the central frequency of the signal, the overall trend of the level value decreases rapidly. The energy distribution diagram of FH signal is shown in Figure 15.
Signal energy detection steps are as follows:

5)
The characteristic value of signal energy is used as input parameter for clustering, and the results are shown in Figure 16, 16-(a) is correspond to Figure 14-(a2), Figure 16-(b) is refer to Figure 14-(b2).

8) SIGNAL DWELL TIME DETECTION
Signal dwell time detection steps are as follows: 1) ∀S ∈ , set the starting time as t s , the ending time as t e , and calculate the dwell time d = t e − t s .
2) Calculate all signals in according to the step 1), and get the signal dwell time set: D = {d 1 , d 2 , · · · , d k }, Diagram of signal dwell time is shown in Figure 17. The results of dwell time clustering for the data processed after energy detection are shown in Figure 18. Three classes with the number of objects satisfying the condition are shown in Figure 19.
It can be seen from the diagram, the objects in their respective class (a), (b) and (c) are extremely similar. However, in the actual electromagnetic environment, there may be many FH signals of the same type of UAV at the same time. Therefore, it is necessary to classify the two types of hop signals according to the time sequence and the time gap stability characteristics.

9) CHRONOLOGICAL SORTING
The FH signals generated by UAV do not overlap in time series of each FH signal, so the possible situation in the detection environment is analyzed as follows: 1) If the time gap between all signals is stable or integer times of the minimum time period, there is only one UAV.
2) If multiple signals are crossed in time and spaced at a certain distance in frequency, two or more UAVs exist.
The characteristics of FH signal time sequence is shown in Figure 20, Frequency hopping signals with the same tag values (such as 1-7) but different background colors are overlapped in time.   3) Process all elements of O until each element is classified, and all signals in these categories do not overlap in time.
4) The categories in which the total number of elements less than 5 are eliminated, and the rest is the detection result of UAV FH signal.
In addition, two UAVs of the same model can be classified into a category, we can continue determine each category whether have two or more uavs by the common-sense knowledge which is that a single uav always can't appear two FH signals at a time, and then get the final results of uav FH signals detection.

C. EXPERIMENTAL RESULTS
The experiment in this paper mainly aims at the situation that there is only one UAV or two types of UAVs coexist. The proposed algorithm is applied to signal detection and verification, and the detection effect is compared with K-means and other methods.

1) ONLY UAV-A EXIST
Turn on the remote control of UAV-A, and use the radio receiver to scan the frequency band of 2.4GHz to monitor the signal. In the experiment, the spectrum (Figure 21, left) and processed signal (Figure 21, right) are presented by waterfall diagram. Different types of signals are distinguished by different colors, and the number of color types represents the number of UAVs. The waterfall map after threshold preprocessing of spectrum data is shown in Figure 21.
The processed results of the improved clustering algorithm, K -means (k = 2), K -means ++ (k = 2), DBSCAN and Multi-hop Auto-correlation algorithm on frequencysweep spectrum data of only existing UAV-A are shown in Figure 22. The results are shown in Figure 22(a-e), in which using K -means the algorithm doesn't detect the signal of UAV. Other methods all show that there is a UAV FH at least.
Through an in-depth analysis, the reason of K -means algorithm can't detect the signal is that the bandwidth (about 1MHz) of FH signal of this type of UAV is extremely close, which makes it unable to be divided into preset K (k = 2) categories. K -means ++ method also has the problem of pre-set K (k = 2), but when selecting the initial cluster center, starting from the second cluster center, the object farthest from the first cluster center is selected as the candidate cluster VOLUME 9, 2021 center, that is, when the clustering elements are extremely close, at least one classification can be obtained, and the result is different from K -means.

2) ONLY UAV-B EXIST
Turn on the UAV-B remote control to test and verify only this UAV activity within the 2.4GHz frequency band. The spectrum data and the results after threshold processed of UAV-B signals are shown in Figure 23, the signal detection results shown in Figure 24.
As shown in Figure 24(a-e), the FH signal of UAV-B in current electromagnetic environment can be detected effectively by using improved clustering algorithm, and only one UAV signal be found. Although K -means (k = 2) algorithm can detect most FH signals, some interference signals cannot be removed in the classification results. Similarly, the detection results of K -means ++ and DBSCAN algorithms contain some interference signals. The detection algorithm of Multi-hop Auto-correlation can retain all the undisturbed FH signals, but the detection results contain a large number of noise signals due to its strong correlation with unknown noise.

3) UAV-A AND UAV-C COEXISTENCE
In order to test the ability of these algorithms to detect FH signals of multiple UAVs, the remote controllers of two UAVs (A and C) are turned on at the same time. The spectrum data threshold preprocessing results are shown in Figure 25.
As shown in Figure 26 (a-e), different colors indicate the types of signals detected (i.e. the number of UAVs). Two types of FH signals are detected by improved clustering and Multi-hop Auto-correlation algorithm. K -means (k = 2) algorithm only detects one type of FH signal of UAV. Although K -means ++ (k = 2) algorithm and DBSCAN detect two types of FH signals of UAV, a large number of noise and interference data need to be processed in the later parameter estimation stage.

V. EXPERIMENTAL ANALYSIS A. ANALYSIS OF RESULTS
The UAV signals under different conditions are detected by improved clustering algorithm, K -means, K -means ++ , DBSCAN and Multi-hop Auto-correlation. The number of detected UAV signals, the number of error detections, accuracy rate, error rate, running time and number of detected UAVs are shown in Table 4. Using the error detection rate and accuracy rate to analyze the performance of each algorithm under different experimental conditions, as shown in Figure 27 and Figure 28. Among these algorithms, the number of detected signals is the number of UAV FH signals by each algorithm, the number of error detection signals is the number of detected signals  that are not UAV FH signals. The accuracy rate is the proportion of the number of detected correct FH signals and the actual number of received FH signals. The error detection rate is the ratio between the number of non-UAV FH signals and the number of signals received from UAV. The running time is the total time from the beginning of algorithm running to the result output. The number of UAVs detected is the number of UAVs determined.
Basically, table 4 shows that algorithm of this paper has a lower error detection rate and accuracy performance VOLUME 9, 2021 advantage when UAV-A FH signal activity only, the densitybased DBSCAN clustering method has the highest accuracy, to some extent, it confirms that UAV FH signal has the advantage of density analysis, and on the contrary the traditional K -means method has a bad accuracy performance. On the one hand, the accuracy rate of the paper algorithm proposed in this paper is higher than most of the other four algorithms, whether there is only one FH signal activity of UAV-A or UAV-B, or when UAV-A and C coexist, indicating that the algorithm has certain stability, as shown in Figure 28. It has significant adaptability to detect UAV FH signals in different environments. On the other hand, the error detection rate of this algorithm is far lower than other algorithms in different cases, as shown in Figure 27, which improve the efficiency for further analysis and processing of UAV signals.
In short, through a variety of different tests, the algorithm in this paper has obvious advantages in signal detection accuracy and error detection rate on FH signal detection, and can be applied to the detection of UAV FH signal in the actual electromagnetic environment.

B. ANALYSIS OF PARAMETERS ESTIMATION
The following parameters are estimated by the results of those algorithms: dwell time, bandwidth, duration and span. As shown in Figure 29 and Figure 30, it is easy to find that when only UAV-A and UAV-B exist, the four parameters are obtained by the algorithm proposed in this paper are closer to the parameters of actual UAV. As shown in Figure 31 and Figure 32, when UAV-A and UAV-C exist at the same time,    the parameters obtained by the algorithm in this paper are also closer to the actual UAV parameters respectively. The detailed data are shown in Table 5, Table 6 and Table 7.
In the parameter estimation of UAV-A, due to the very short dwell time characteristics and maybe the limited amount of sampling data, it is impossible to detect and estimate the signal data of UAV-A using K -means.
When UAV-A and UAV-C exist at the same time, the parameter estimation of UAV-A after DBSCAN algorithm   is almost consistent with the actual parameters, which play a good performance. But when estimating the parameters of UAV-C, the proposed algorithm in this paper has a better estimation characteristic. In summary, the algorithm in this paper has a good and stable ability for FH signal analysis.

VI. CONCLUSION
In this paper, through adaptive noise threshold calculation, FH signal waveform characteristics and peak characteristics are developed. In different experimental environments, according to the pre-designed UAV signal detection process, the improved clustering analysis algorithm is used to detect, analyze and estimate the parameters of UAV FH signal in electromagnetic environment with low SNR. Compared with K -means, K -means ++ , DBSCAN and Multi-hop Auto-correlation algorithm, the experimental results show that the proposed method can be used to quickly and effectively detect the FH signal of UAV, in addition, the estimated parameters are closer to the actual value. It seems that the proposed method is useful and alternative tool to study UAV signal detection and real-time tracking.
In experimental analysis, the proposed method and algorithm are suitable for detection and parameter estimation of the FH signal. Our future works will be focused on UAV signal source location and tracking.