TDMA signal sorting method based on DBSCAN clustering

Abstract

The invention proposes a TDMA signal sorting method based on DBSCAN clustering, and relates to the technical field of signal sorting, and the method comprehensively considers the influence of various transmitter structures on TDMA signal features, evaluates the recognition precision and implementation complexity of different features, optimizes the burst power, carrier frequency offset and RMS EVM features, and improves the recognition precision and implementation complexity of TDMA signals. The method comprises the following steps: firstly, carrying out related feature extraction on a collected TDMA signal, and then carrying out clustering operation by using feature data as input and a DBSCAN algorithm according to a distance relationship between data points and a density condition. According to the distribution condition of the feature data points, the signals with similar features are clustered into one class, that is, the signals of the same user are clustered together, so that accurate user separation is realized. Compared with a two-dimensional feature clustering algorithm, the method is based on three-dimensional feature clustering, the calculation complexity is improved, meanwhile, the sorting accuracy is improved, and the problem that correct sorting cannot be completed only through frequency deviation and burst power is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal sorting, and particularly to a TDMA signal sorting method based on DBSCAN clustering. Background Art

[0002] In today's complex, changeable and electromagnetic interference-filled information environment, various signals exhibit characteristics of high density, diversity and dynamic change. Whether it is an electronic warfare system in the military field or the development of communication technology in the civilian field, how to accurately sort and extract target communication signals from numerous mixed signals to achieve efficient spectrum management and reliable communication quality assurance, signal sorting plays a crucial role in extracting valuable information and ensuring the safe and stable operation of the system.

[0003] Early research on signal sorting mainly focused on methods based on traditional parameter features. Yuan Jiangyou et al. aimed at the characteristics of multiple pulses in a single user time slot of a TDMA system, used Hilbert transform to highlight the pulse edge differences, and extracted the subtle features of the pulses by statistical smoothing method to sort users; Luo Wuzhong et al. also aimed at the time slot allocation of the TDMA system and the continuity of the time difference of user arrivals, and proposed to sort users by measuring the arrival time difference, which required using multiple antennas to receive signals, and the system implementation was complex. In addition, Wang Ming et al. based on the physical layer characteristics of TDMA signals and the Doppler frequency offset caused by the relative motion between the satellite receiving station and the satellite, performed burst signal detection on the signals, and extracted two features of burst power and carrier frequency offset in the time-frequency domain for signal sorting.

[0004] However, with the rapid development of communication technology, the structure of TDMA transmitters shows a trend of diversification, from direct orthogonal up-conversion transmitters to digital intermediate frequency transmitters, and then to multi-stage mixing transmitter structures; the application of adaptive coding modulation technology and link closed-loop control in TDMA systems makes the feature differences of signals become more and more refined, and the accuracy and reliability of traditional sorting methods for signal sorting face severe challenges. Summary of the Invention

[0005] Aiming at the defects and deficiencies existing in the above-mentioned prior art, comprehensively considering the influence of various transmitter structures on the characteristics of TDMA signals, evaluating the recognition accuracy and implementation complexity of different features, and preferring burst power, carrier frequency offset and RMSEVM features, the present invention proposes a TDMA signal sorting method based on DBSCAN clustering, and the method includes the following steps:

[0006] Step 1: Intercept TDMA signals of the same carrier, and use the double-sliding window energy detection method to judge the position of the time slot of different user signals;

[0007] Step 2: Determine the start and end positions of each time slot, and separate the signals of each time slot;

[0008] Step 3: Use the frequency offset estimation method based on particle swarm optimization to extract the features of the signals in each separated time slot. The features include carrier frequency offset, burst power, and RMS EVM value;

[0009] Step 4: Organize the data to form a three-dimensional data set P = {p 1 , p 2 ,..., p i} of carrier frequency offset, burst power, and RMS EVM value;

[0010] Step 5: Classify the set using the DBSCAN clustering algorithm, and optimize the clustering effect by adjusting the two key parameters of the Eps neighborhood and MinPts;

[0011] Step 6: Output the final sorting result and determine the number of users existing in the TDMA signal of the carrier.

[0012] Preferably, the extraction process of the carrier frequency offset feature in Step 3 is as follows:

[0013] Step 3.1: Determine the sample points of the time slot signal to be estimated, construct the objective function, and set the particle swarm search interval as [x min , x max ; Step 3.2: Initialize the velocity and position of the particles, substitute the individual optimal solutions of each particle at this time into the constructed objective function to obtain the fitness values of the particles, and take the solution with the largest fitness value as the global optimal solution g 0 ;

[0014] Step 3.3: Perform the following operations in the i-th iteration:

[0015] a) Update the particle velocity;

[0016] b) Update the particle position;

[0017] c) Update the individual optimal solution of the particle;

[0018] d) Update the global optimal solution g i ;

[0019] Judge whether the maximum number of iterations is reached. If not, go to Step 3.3. Otherwise, terminate the iteration and output the global optimal solution g end , which is the frequency offset estimation value.

[0020] Preferably, the extraction method of the burst power feature in Step 3 is: calculate the short-term root mean square of each time slot to obtain the burst power feature value:

[0021]

[0022] Preferably, the extraction method of the RMS EVM value feature in step 3 is: The RMS EVM is defined as the square root of the ratio of the average error vector power P error to the average reference power P reference :

[0023]

[0024] Preferably, the specific process of step 5 is as follows:

[0025] Step 5.1: Cluster the set P using the DBSCAN clustering algorithm to determine the two parameters of the Eps neighborhood and MinPts;

[0026] Step 5.2: For each point p in the set P, calculate all the points within its neighborhood with a radius of Eps in three-dimensional space, and these points form the neighborhood N Eps (p) of point p;

[0027] Step 5.3: Determine the core points. Traverse all the points in the dataset. For each point p, count the number of points within its neighborhood N Eps (p). If the number of points in N Eps (p) is greater than or equal to MinPts, then mark point p as a core point; otherwise, mark it as a non-core point;

[0028] Step 5.4: Select an unprocessed core point p from the dataset as the starting point, create a new cluster C, and add point p to cluster C;

[0029] Step 5.5: For each point q within the neighborhood N Eps (p) of point p, if q is a core point, then also add it to cluster C, and add the points in the neighborhood N Eps (q) of q that have not been added to cluster C;

[0030] Step 5.6: Repeat the above steps to continuously expand cluster C until no new points can be added to cluster C, and at this time, the division of one cluster is completed.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] Compared with the clustering algorithm of two-dimensional features, the present invention is based on three-dimensional feature clustering. While increasing the computational complexity, it also improves the correct sorting rate and effectively solves the problem that correct sorting cannot be completed solely by frequency offset and burst power. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the QPSK modulation domain error.

[0034] Figure 2 This is the algorithm flowchart of the present invention.

[0035] Figure 3 This is the schematic diagram of the density-reachable generated clustering clusters of the present invention.

[0036] Figure 4 This is the convergence curve graph of the algorithm of the present invention with the number of iterations.

[0037] Figure 5 This is the estimation variance of PSO and other algorithms at different signal-to-noise ratios.

[0038] Figure 6 This is the unbiased estimation range analysis of PSO and other algorithms.

[0039] Figure 7 a) Signal constellation diagram of QPSK and the corresponding mean square error vector magnitude value

[0040] Figure 7 b) Signal constellation diagram of QAM and the corresponding mean square error vector magnitude value

[0041] Figure 8 This is the comparison graph of the sorting correct rate. Specific implementation manner

[0042] The following further explains the present invention in conjunction with the accompanying drawings and specific embodiments.

[0043] Considering the diversity of transmitter structures and the application of new technologies such as adaptive coding modulation and link closed-loop control in TDMA systems, it will have a certain impact on the generation and calculation of these features. Therefore, it is first necessary to optimize the TDMA signal features. The following lists some typical features of TDMA signal sorting:

[0044] ① Frequency offset

[0045] Due to reasons such as the local oscillator frequency of the transmitter and the possible Doppler effect, there will be differences in frequency offset in the received signal. Using this feature, the frequency offset feature can be optimized to distinguish the signals of different users. For example, in the GSM system, the frequency error requirement is relatively strict, generally requiring the frequency offset to be around ±0.1 ppm to ±0.5 ppm.

[0046] Traditional carrier frequency feature estimation based on chirp-Z transform (CZT) spectrum refinement analysis can find the main frequency components of a signal in the frequency domain by performing a CZT transform on the received signal, thereby estimating the carrier frequency. However, to achieve the estimation accuracy required for signal sorting, the computational complexity of this algorithm is extremely high. Therefore, a frequency offset estimation method based on particle swarm optimization is introduced here. This algorithm uses the likelihood function of frequency offset estimation as the objective function and simulates swarm intelligence to search for the optimal solution. At the same time, the carrier frequency offset also has high adaptability to different transmitters and will not be seriously affected by changes in the transmitter structure, so the accuracy of the algorithm and the distinguishability of features are sufficient to support it as an excellent sorting feature.

[0047] ② Burst power

[0048] The physical meaning of burst power is the energy intensity of a signal in a short period of time. In a TDMA system, signals in different time slots come from different users or communication sources, and their transmission powers usually vary, resulting in differences in the burst powers of signals in different time slots. For the quantization calculation of burst power, we usually use the short-term root mean square method. Specifically, this refers to obtaining the maximum value of the signal within a precisely defined short time window range and determining its minimum value at the same time. The difference between the two is the burst power value calculated by the short-term root mean square:

[0049]

[0050] In actual engineering application scenarios, the estimation accuracy of burst power can be controlled within 1 dB. This accuracy level is achieved through a large number of experimental optimizations and algorithm calibrations, indicating that our capture of signal burst power has reached a relatively high level of accuracy. However, it cannot be ignored that the power control link in the system will bring difficult problems to this characteristic of burst power. After the system performs power control operations, the distinct burst power distinguishability of signals in different time slots will be greatly reduced. Signals from different users or communication sources that could originally be clearly distinguished by burst power differences now become blurred and difficult to distinguish.

[0051] Therefore, in order to more accurately achieve the signal sorting task and achieve efficient communication, we urgently need to organically combine this key feature of burst power with other distinguishable features to effectively cope with complex and changing communication situations.

[0052] ③ Symbol rate

[0053] In a communication link, even when using communication devices of the same model, including terminal devices such as modems and upconverters. Due to factors such as the operating state of the devices, the operating environment, and the differences between chips and components, slight differences in symbol rate and carrier frequency objectively exist. Currently, the research on symbol rate estimation mainly focuses on four aspects, namely delay multiplication, cyclic spectrum method, wavelet transform method, and deep learning method.

[0054] Although differences in symbol rate parameters always exist, in a TDMA system, the symbol rates are basically the same, and the device manufacturer's protocol stipulates that the maximum allowable error in the terminal rate accuracy is 10-6Rs. At this time, the symbol rate estimation accuracy of relevant algorithms cannot reach the level of distinguishing their differences, so the symbol rate cannot be used as a preferred feature to participate in the sorting of TDMA signals.

[0055] ④ Pulse front edge rise time

[0056] Compared with an ideal radiation source, there are many non-linear characteristics of devices in an actual radiation source. This difference will cause small time-varying fluctuations in the signal envelope, and these small fluctuations are reflected in the signal as phenomena such as changes in the rising and falling edges of the signal.

[0057] The generation of pulse front edge envelope characteristics is mainly related to the relationship between the time domain and frequency domain characteristics of the signal, especially the changes in the rising or falling edge of a non-ideal pulse signal. However, the computational complexity of pulse front edge characteristics is extremely high, and it is greatly affected by noise, so pulse front edge characteristics are also not suitable for the proposed TDMA signal sorting.

[0058] ⑤ Carrier leakage

[0059] Carrier leakage occurs because of the non-ideal characteristics of analog circuit components such as amplifiers and mixers at the transmitter and receiver ends of a communication system. These non-ideal characteristics may cause the local oscillator signal to leak to the output end, thus generating carrier leakage. However, carrier leakage only has obvious characteristics in a direct quadrature upconversion transmitter, and its adaptability to different transmitter structures is too weak, so it is also not suitable for TDMA signal sorting.

[0060] ⑥ RMS EVM value

[0061] High-frequency links generally use adaptive modulation methods. Taking DVB-S2 as an example, DVB-S2 can dynamically adjust the modulation method and coding rate according to changes in channel conditions. In the process of TDMA signal sorting, the selection of modulation method-related features is of great significance. TDMA signals under different modulation styles often exhibit unique distinguishable features. Among them, modulation method recognition is relatively difficult, and usually, it is equivalent to power during research.

[0062] Further in-depth exploration reveals that the characteristics based on modulation methods do not exist in isolation. They are fully manifested in the RF fingerprint characteristics generated by the device, which can be reflected by the calculated Error Vector Magnitude (EVM). EVM can be used to characterize the combined influence of multiple fingerprint sources on the final signal, including DC bias, phase noise, I / Q quadrature offset, I / Q gain imbalance, and other RF fingerprint characteristics. All these RF fingerprint characteristics are intertwined and fused together through this key indicator of EVM, jointly shaping each unique signal characteristic, such as Figure 1 as shown.

[0063] The Error Vector (EV) is the vector difference between the actual transmitted signal P means and the ideal signal P ref , and EVM is the ratio of the magnitude of the error vector to the magnitude of the reference signal:

[0064]

[0065] EVM applies to each transmitted and received symbol. Finally, the modulation method and the EVM value are normalized, and RMSEVM is defined as the square root of the ratio of the average error vector power P error to the average reference power P reference :

[0066]

[0067] Specially, in the actual communication scenario, there is also a very challenging situation. The same user does not always adopt a fixed modulation method during communication and sometimes changes the modulation method based on various factors. In the research scope explored in the present invention, since once the modulation method changes, the signal characteristics will change significantly and the signal differences corresponding to different modulation methods are extremely large, the situations before and after the modulation method change generated by the same user are regarded as the signals of two different users.

[0068] Taking into comprehensive consideration the discrimination degree, calculation complexity of the above characteristics, and the adaptability to the transmitter structure, the following preferred characteristics are finally adopted to improve the sorting accuracy, as shown in Table 1.

[0069] Table 1 Finally adopted preferred characteristics

[0070]

[0071] Such as Figure 2As shown, the TDMA signal sorting method based on DBSCAN clustering has the following general workflow: First, relevant features of the collected TDMA signals are extracted, and then, using these feature data as input, the DBSCAN algorithm performs clustering operations based on the distance relationship and density among data points. According to the distribution of these feature data points, signals with similar features are clustered into one category, that is, signals of the same user are grouped together, thereby achieving accurate user separation.

[0072] As Figure 3 shown, based on the radius Eps, the data is divided into the Eps neighborhood sets of each point. The so-called Eps neighborhood refers to the set composed of all points whose distance from a certain point p is less than or equal to Eps. At the same time, the number of data points MinPts is defined. Within the range of the radius Eps, if the number of points exceeds MinPts, these points are defined as core points; if the number of points is less than MinPts but within the neighborhood of a core point, these points are defined as border points; the remaining points are noise points. This algorithm searches for clusters by checking the Eps neighborhoods of each point in the dataset. If the number of points contained in the Eps neighborhood of point p exceeds MinPts, a cluster with p as the core object will be created. Starting from this core object, DBSCAN will iteratively aggregate samples that are directly density-reachable from these core objects, thereby generating clustering clusters until all core objects have been visited.

[0073] In the DBSCAN algorithm, the two key parameters, the Eps neighborhood and MinPts, have a decisive impact on the clustering results. Among them, the Eps neighborhood defines the set range of points whose distance from a certain point is less than or equal to a specific value. This parameter dominates the scale and shape of the clustering clusters. When the value of Eps is relatively small, many data points that should be grouped into the same cluster will be split into different clusters, resulting in an overly fragmented clustering result. On the contrary, if the value of Eps is relatively large, signals of different users may be grouped into the same cluster, and it is impossible to accurately distinguish signals of different users.

[0074] The MinPts parameter sets the minimum number of data points that a point should contain to be judged as a core point within the Eps neighborhood. If the value of MinPts is too large, some neighborhood point data that originally belonged to the clustering cluster will be misjudged as noise, resulting in the omission of some signals; if the value of MinPts is too small, the number of core points will be too large, and noise may be misrecognized as data within the cluster, interfering with the accuracy of clustering and the subsequent signal sorting effect. Only through reasonable parameter setting can the DBSCAN algorithm achieve optimal performance in the TDMA signal sorting process and improve the sorting accuracy.

[0075] Simulation experiment:

[0076] 1. Optimal feature preprocessing

[0077] The carrier frequency offset is estimated using the particle swarm optimization algorithm. The corresponding simulation conditions are set as follows: QPSK modulation is used, the number of symbols is 64, the normalized frequency offset is 0.1, the number of particles is 6, the inertia factor c0 = 0.4, the learning factors c1 = c2 = 1, and 10,000 Monte Carlo simulations are performed. The convergence curve is as Figure 4 shown. For QPSK, the algorithm converges to near the true value after 7 iterations, and the convergence speed increases with the increase in signal-to-noise ratio.

[0078] The study found that at a signal-to-noise ratio of 10 dB, the relative error obtained is 0.4%. As the signal-to-noise ratio increases to 20 dB, the estimation accuracy is significantly improved, and the relative error almost approaches 0. This indicates that in a high signal-to-noise ratio environment, the estimation of signal parameters becomes more accurate.

[0079] Next, the estimation variances of the particle swarm optimization algorithm (PSO) and several common algorithms such as FFT, M&M, and Fitz are compared. Variance is a statistic used to measure the degree of data dispersion. In signal parameter estimation, the smaller the estimation variance, the more stable the result of the algorithm estimation and the smaller the fluctuation. Figure 5 The comparison of the estimation variances of these algorithms under different signal-to-noise ratio conditions is shown. It can be intuitively seen from the figure the changing trends of the estimation variances of different algorithms with the change in signal-to-noise ratio. At the same time, Figure 6 the unbiased estimation ranges of multiple algorithms are presented. The larger the unbiased estimation range, the more accurate the estimation can be given by the algorithm under a wider range of conditions. Through the analysis of these two figures, it can be concluded that the particle swarm optimization algorithm performs excellently in terms of performance and adaptability. It can maintain a relatively small estimation variance and a large unbiased estimation range under different signal-to-noise ratio conditions, ultimately proving that the frequency offset estimated by this algorithm can meet the estimation accuracy and discrimination required for signal sorting. This means that in practical applications, when signal sorting is required, the particle swarm optimization algorithm can accurately estimate the frequency offset of the signal and effectively distinguish different signals.

[0080] Figure 7 The signal constellation diagrams and the corresponding root mean square error vector magnitude (RMS EVM) values under two different modulation methods, QPSK and 16QAM, are shown. By setting different radio frequency fingerprint parameters, the finally obtained RMS EVM values will be significantly different. And the simulation results show that at a signal-to-noise ratio of 10 dB, the RMS EVM values under the two modulation methods of QPSK and 16QAM have good discrimination.

[0081] 2. Data verification of the clustering sorting effect

[0082] The simulation experiment conditions are set as follows: the signal modulation methods are QPSK and 16QAM, there are 7 users, the user time slot length is about 0.586 ms, the sampling rate fs = 5 MHz, the symbol rate fd = 1.024 MHz, the carrier frequency fc = 2 MHz. Based on the frequency offset and power difference, different RF fingerprint features are added to the signal to distinguish the RMS EVM values of different user signals.

[0083] The experiment conducts sorting tests on the two-dimensional data regarding the frequency offset and power, and the three-dimensional data after introducing the RMS EVM value respectively. The test results show that, also under the condition of signal-to-noise ratio of 5 - 15 dB, the correct sorting rate of the three-dimensional clustering of the TDMA burst signal containing 7 users has an obvious improvement of 1 user, as Figure 8 shown.

[0084] In the modern communication environment, there are various electromagnetic interference sources. These interference signals are superimposed on the TDMA signal in aspects such as time domain and frequency domain, masking or distorting the TDMA signal characteristics and increasing the difficulty of accurate extraction and sorting. Facing these challenges, on the one hand, the signal preprocessing algorithm can be optimized, and more advanced filtering, noise reduction and other technologies can be adopted to reduce the influence of interference signals as much as possible and enhance the stability of the extracted features; on the other hand, the feature dimension can be increased, comprehensively considering the feature information in aspects such as time domain, frequency domain, modulation method, etc., to construct a more comprehensive and accurate feature description, so as to improve the sorting performance of TDMA signals in complex environments. At the same time, the sorting algorithm is continuously improved to make it have better adaptability and be able to handle dynamic changes such as the number of users and time slot allocation, so as to achieve more reliable TDMA signal sorting.

[0085] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A TDMA signal sorting method based on DBSCAN clustering, characterized in that: The steps include: Step 1: intercept the TDMA signal of the same carrier and use the double sliding window energy detection method to determine the time slot location of different user signals; Step 2: Determine the start and end positions of each time slot and separate the signal of each time slot; Step 3: Extract features of the signal of each time slot after separation using a frequency deviation estimation method based on particle swarm optimization, wherein the features include carrier frequency deviation, burst power, and RMS EVM value; Step 4: Arrange the data to form a three-dimensional data set P = {p1, p2, ..., p i }; Step 5: Use the DBSCAN clustering algorithm to classify the set, and adjust the two key parameters of Eps field and MinPts to optimize the clustering effect; Step 6: Output the final sorting result to determine the number of users existing in the TDMA signal of the carrier.

2. The TDMA signal sorting method based on DBSCAN clustering according to claim 1, characterized in that: The extraction process of the carrier frequency deviation feature in step 3 is as follows: Step 3.1: Determine the time slot signal sample points to be estimated, construct the objective function, and set the particle swarm search interval to [x min ,x max ]; Step 3.2: Initialize the speed and position of the particle, substitute the individual optimal solution of each particle into the constructed objective function to obtain the fitness value of the particle, and take the solution with the largest fitness value as the group optimal solution g0; Step 3.3: Iterate i times and perform the following operations: a) Update particle velocity; b) Update particle position; c) Update the individual optimal solution of the particle; d) Update the group optimal solution g i ; Determine whether the maximum number of iterations has been reached. If not, proceed to step 3.

3. Otherwise, terminate the iteration and output the optimal solution g end , which is the estimated value of frequency offset.

3. The TDMA signal sorting method based on DBSCAN clustering according to claim 1, characterized in that: The method for extracting the burst power feature in step 3 is: calculating the short-time root mean square of each time slot to obtain the burst power feature value:

4. The TDMA signal sorting method based on DBSCAN clustering according to claim 1, characterized in that: The method for extracting the RMS EVM value feature in step 3 is as follows: RMS EVM is defined as the average error vector power P error With the average reference power P reference The square root of the ratio of :

5. The TDMA signal sorting method based on DBSCAN clustering according to claim 1, characterized in that: The specific process of step 5 is as follows: Step 5.1: Use the DBSCAN clustering algorithm to cluster the set P and determine the two parameters of Eps area and MinPts; Step 5.2: For each point p in the set P, calculate all the points in its neighborhood with Eps as radius in the three-dimensional space. These points constitute the neighborhood N of point p. Eps (p); Step 5.3: Determine the core point, traverse all points in the data set, and for each point p, count its neighborhood N Eps (p) points, if N Eps If the number of points in (p) is greater than or equal to MinPts, then point p is marked as a core point; Otherwise, mark it as a non-core point; Step 5.4: Select an unprocessed core point p from the data set as the starting point, create a new cluster C, and add point p to cluster C; Step 5.5: For the neighborhood N of point p Eps For each point q in (p), if q is a core point, it will also be added to cluster C, and the neighborhood N of q will be Eps The points in (q) that have not yet been added to cluster C are also added; Step 5.6: Repeat the above steps and continue to expand cluster C until no new points can be added to cluster C. At this time, the division of a cluster is completed.

Images