A method for identifying radio frequency signals of unmanned aerial vehicles based on Tsetlin Machine

Abstract

This invention discloses a UAV RF signal recognition method based on Tsetlin Machine. First, the UAV's RF I / Q signals are sampled and denoised using Short Time Fourier Transform (STFT) and mathematical morphological filtering. Then, through time-frequency projection compression and first-order alignment, the complex two-dimensional time-frequency graph is reconstructed into a simplified one-dimensional Boolean feature vector incorporating expert prior statistical features. Finally, a lightweight pure logic recognition model based on Tsetlin Machine™ directly receives this one-dimensional Boolean feature vector as input and performs discrete reinforcement learning training on the software side. This invention also has the following advantages: 1. Efficient dimensionality reduction, preserving core physical features, effectively solving the problem of feature aliasing between similar UAV types. 2. Avoiding high-precision floating-point operations at the data level, replacing the complex quantization, compilation processes, and DSP multiplier operations required by traditional deep learning neural networks with pure logic gate circuits.

Description

Technical Field

[0001] This invention belongs to the field of UAV detection and communication signal processing technology. More specifically, it relates to a UAV radio frequency signal identification method based on Tsetlin Machine, which is used for lightweight identification of UAV radio frequency signals in complex electromagnetic environments, and is particularly suitable for edge device deployment with limited computing power and power consumption. Background Technology

[0002] In recent years, the widespread use of micro and small drones has brought serious challenges to low-altitude airspace security. In anti-drone detection systems, passive detection technology based on radio frequency (RF) signals has advantages such as long detection range and good concealment.

[0003] However, existing UAV radio frequency signal recognition methods have significant limitations in complex electromagnetic environments: traditional time-frequency domain statistical feature methods have weak anti-interference capabilities and are prone to inter-class feature aliasing when facing UAV models using similar underlying protocols, resulting in multi-class accuracy hitting a physical limit. On the other hand, while deep learning-based solutions, such as Convolutional Neural Networks (CNNs), can automatically mine high-dimensional implicit features, they require a large amount of expensive digital signal processing hardware resources due to the massive floating-point multiplication and addition operations involved. This makes it difficult to efficiently deploy such models on edge devices with limited power consumption and computing power (such as portable detection devices and FPGAs), creating a contradiction between high recognition accuracy and high hardware deployment costs. Summary of the Invention

[0004] To address the contradiction between high recognition accuracy and high hardware deployment costs in existing technologies, this invention provides a UAV radio frequency signal recognition method based on Tsetlin Machine. The method aims to achieve low-power, low-latency, and high-precision edge hardware deployment by compressing the logical features of radio frequency signals and classifying them using pure Boolean logic.

[0005] To achieve the above-mentioned objectives, the present invention provides a UAV radio frequency signal identification method based on the Tsetlin Machine, characterized by comprising the following steps:

[0006] (1) Radio frequency signal preprocessing to generate a two-dimensional time-frequency diagram

[0007] First, the radio frequency I / Q signals of the drone are sampled to obtain the radio frequency sampled signal. And generate a spectrum using short-time Fourier transform. :

[0008]

[0009] in, It is a real-valued symmetric window function. For the sliding step index on the time axis, Angular frequency;

[0010] Then, by calculating the spectrum The square of the modulus, i.e. the power spectral density, yields the two-dimensional time-frequency diagram of the UAV's radio frequency signal. :

[0011]

[0012] (2) Generate the initial binary logic matrix

[0013] First, from the two-dimensional time-frequency diagram Random selection Pixels, and calculate The intensity values ​​of the nth pixel are then sorted from low to high, and the nth pixel is selected. The intensity value of each pixel is used as the threshold. , two-dimensional time-frequency graph Binarization yields the initial binary logical matrix. :

[0014]

[0015] (3) Noise reduction processing using mathematical morphology filtering

[0016] For the initial binary logic matrix Perform mathematical morphological filtering to compress it to Standardized logic matrix ,in, Indicates row and column index;

[0017] (4) Perform maximum value projection

[0018] Standardized logic matrix Transform into time-domain and frequency-domain sequences

[0019] 4.1) Time-domain projection to obtain the time-domain sequence. :

[0020]

[0021] 4.2) Frequency domain projection to obtain the frequency domain sequence. :

[0022]

[0023] (5) First-end alignment to eliminate time offset

[0024] For time-domain sequences ,from Start towards Perform a search at the index where the first logical "1" appears, and denote it as... And logically shift the vector to the left. The bits are used to obtain the aligned time-domain vector. :

[0025]

[0026] (6) Add statistical feature information

[0027] 6.1) Extracting the frequency domain sequence The end Invalidate all zeros to obtain the time-domain vector. ;

[0028] 6.2) Constructing using expert prior statistical features The bit sequence E, where the expert prior statistical features are the number of frequency hopping blocks in the two-dimensional time-frequency diagram, whether there is a large-area image transmission signal in the energy density, and spatial distribution information;

[0029] 6.3) Generate a fixed-length one-dimensional Boolean eigenvector X:

[0030]

[0031] (7) Construct a lightweight pure logic recognition model for recognition.

[0032] The Tsetlin Machine (TM) is used as the core classifier to classify the one-dimensional Boolean feature vector X to obtain the category of the UAV radio frequency signal.

[0033] The objective of this invention is achieved as follows:

[0034] The present invention provides a UAV radio frequency signal recognition method based on Tsetlin Machine. First, the UAV radio frequency I / Q signal is sampled and then denoised using short-time Fourier transform (STFT) and mathematical morphological filtering. Then, through time-frequency projection compression and first-order alignment, the complex two-dimensional time-frequency graph is reconstructed into a simplified one-dimensional Boolean feature vector that incorporates expert prior statistical features. Finally, based on the lightweight pure logic recognition model of Tsetlin Machine™, the one-dimensional Boolean feature vector is directly received as input, and discrete reinforcement learning training is completed on the software side.

[0035] In addition, the UAV radio frequency signal identification method based on Tsetlin Machine of the present invention also has the following beneficial effects:

[0036] (1) Efficient dimensionality reduction and preservation of core physical features: The feature extraction method of the present invention effectively preserves the key frequency points, dwell time and duty cycle of UAV frequency hopping spread spectrum (FHSS) and orthogonal frequency division multiplexing (OFDM) signals. Furthermore, by introducing expert prior statistical feature bits, it compensates for the information loss after compression and effectively solves the problem of feature aliasing between similar aircraft types.

[0037] (2) This invention avoids high-precision floating-point operations at the data level and replaces the complex quantization, compilation process and DSP multiplier operations required by traditional deep learning neural networks with pure logic gate circuits. Attached Figure Description

[0038] Figure 1 This is a flowchart of a specific implementation of the UAV radio frequency signal identification method based on Tsetlin Machine of the present invention;

[0039] Figure 2 It is a two-dimensional time-frequency diagram generated by preprocessing the radio frequency I / Q signal data of the UAV;

[0040] Figure 3 It is a process of binarization, morphological filtering and compression of a two-dimensional time-frequency graph. Among them, (a) is the power spectrum, i.e., the two-dimensional time-frequency graph, (b) is the initial binarization logic matrix obtained by initial binarization, (c) is the binarization logic matrix obtained by mathematical morphological filtering, and (d) is the normalized logic matrix obtained after compression.

[0041] Figure 4 It is a Boolean heatmap of misaligned recombination of time-domain sequences;

[0042] Figure 5 It is a Boolean heatmap after time-domain sequence alignment and recombination;

[0043] Figure 6 It is a graph showing the recognition rate after 50 rounds of training and validation;

[0044] Figure 7 It is the confusion matrix of the 1D-TM model. Detailed Implementation

[0045] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0046] Figure 1 This is a flowchart of a specific implementation of the UAV radio frequency signal identification method based on Tsetlin Machine of the present invention.

[0047] In this embodiment, as Figure 1 As shown, the UAV radio frequency signal identification method based on Tsetlin Machine of the present invention includes the following steps:

[0048] Step S1: Radio frequency signal preprocessing to generate a two-dimensional time-frequency diagram

[0049] First, the radio frequency I / Q signals of the drone are acquired and sampled to obtain the radio frequency sampled signal. And the spectrum is generated using the short-time Fourier transform (STFT). :

[0050]

[0051] in, These are real-valued symmetric window functions, such as the Hamming window and the Heining window. For the sliding step index on the time axis, ω is the angular frequency.

[0052] Then, by calculating the spectrum The square of the modulus, i.e. the power spectral density, yields the two-dimensional time-frequency diagram of the UAV's radio frequency signal. :

[0053]

[0054] Figure 2 It is a two-dimensional time-frequency diagram generated by preprocessing the UAV's radio frequency I / Q signal data. From Figure 2 We can see that the two-dimensional time-frequency diagram shows the time-frequency distribution characteristics of the UAV frequency-hopping signal: the frequency of the frequency-hopping signal jumps regularly with time.

[0055] Step S2: Generate the initial binary logic matrix

[0056] First, from the two-dimensional time-frequency diagram Random selection Pixels, and calculate The intensity values ​​of the nth pixel are then sorted from low to high, and the nth pixel is selected. The intensity value of each pixel is used as the threshold. , two-dimensional time-frequency graph Binarization yields the initial binary logical matrix. :

[0057]

[0058] This invention employs an adaptive dynamic thresholding algorithm to obtain the threshold used for binarization. In this embodiment, , That is, the intensity value of the 97th percentile (97,000th pixel) sorted from low to high intensity values ​​is selected as the threshold. Using this threshold, the two-dimensional time-frequency graph Binarization:

[0059]

[0060] The threshold As a dynamic threshold, it can robustly separate the energy-concentrated effective communication pulses (FHSS frequency hopping blocks and OFDM image transmission stripes) from complex background noise and generate an initial binary logic matrix.

[0061] In this embodiment, as Figure 3 (b) shows the initial binarization, which yields the initial binarized logic matrix. .

[0062] Step S3: Denoising processing using mathematical morphological filtering

[0063] For the initial binary logic matrix Perform mathematical morphological filtering to compress (normalize) it. Standardized logical matrix ,in, Indicates the row and column index.

[0064] In this embodiment, Figure 3 This demonstrates the process of binarization, morphological filtering, and compression (normalization) of the time-frequency graph. The logistic matrix after morphological filtering is shown below. Figure 3 As shown in (c), the final standardized logic matrix is ​​obtained, as follows: Figure 3 As shown in (d), the standardized logic matrix retains the key features of the original time-frequency diagram.

[0065] Step S4: Perform maximum value projection

[0066] Standardized logic matrix It is converted into a time-domain sequence and a frequency-domain sequence.

[0067] Step S4.1: Time-domain projection to obtain the time-domain sequence. :

[0068]

[0069] That is, the first If any value in the column is 1, then It is 1 if it is true, otherwise it is 0;

[0070] Step S4.2: Frequency domain projection to obtain the frequency domain sequence. :

[0071]

[0072] That is, the first If any row contains a value of 1, then It is 1 if it is true, otherwise it is 0;

[0073] In this embodiment, Then, the time-domain projection yields the time-domain sequence. :

[0074]

[0075] Frequency domain projection yields the frequency domain sequence. :

[0076]

[0077] Step S5: First-end alignment to eliminate time offset

[0078] For time-domain sequences ,from Start towards Perform a search at the index where the first logical "1" appears, and denote it as... And logically shift the vector to the left. The bits are used to obtain the aligned time-domain vector. :

[0079]

[0080] In this embodiment, the visible time-domain vector It is 64.

[0081] In this embodiment, the first five types of UAV radio frequency signal datasets were selected for comparative experiments. Figure 4 , 5 Boolean heatmaps are shown for both unaligned and aligned reassembled versions. In the direct flattening scheme without first-order alignment, the heatmaps of the Phantom 4 Pro (T0010) and MATRICE 200 (T0011) drones are extremely similar, making them difficult to distinguish. In the aligned and reassembled 128-bit feature space, the activation patterns of different protocol models show significant differences.

[0082] Step S6: Add statistical feature information

[0083] Since changing the length of the Boolean feature vector will alter the bus width in subsequent hardware implementations, this invention will retain the total number of bits unchanged and extract the frequency / time domain data. Invalid all-zero bits are used to incorporate expert prior statistical features. Specifically, this includes the following steps:

[0084] Step S6.1: Extract the frequency domain sequence The end Invalidate all zeros to obtain the time-domain vector. In this embodiment, the frequency domain sequence The length is 64. Thus, time-domain vector The length is 59.

[0085] Step S6.2: Construct using expert prior statistical features The bit sequence E, where the expert prior statistical features are the number of frequency hopping blocks in the two-dimensional time-frequency diagram, whether there is a large-area image transmission signal in the energy density, and spatial distribution information;

[0086] Step S6.3: Generate a fixed-length one-dimensional Boolean feature vector X:

[0087]

[0088] In this embodiment, generating a fixed-length one-dimensional Boolean feature vector X can also be expressed as:

[0089]

[0090] The superscript indicates the number of digits.

[0091] Thus, the expert prior statistical features correspond to bits 124-128 of the one-dimensional Boolean eigenvector X, and their specific meanings are as follows:

[0092] 124 bits are frequency hopping active bits: the number of independent frequency hopping blocks (area < 50 pixels) is counted through connected component analysis to characterize the frequency hopping density;

[0093] 125 bits are energy density bits: used to determine whether there is a giant connected region with an area greater than 60 pixels, serving as the physical basis for distinguishing broadband OFDM image transmission signals from narrowband FHSS control signals;

[0094] Bits 126–128 are spatial preference bits: to determine whether the frequency hopping points are mainly concentrated in the upper half or lower half of the observed frequency band, in order to capture the frequency hopping strategy preferences of different manufacturers.

[0095] Step S7: Construct a lightweight pure logic recognition model for recognition.

[0096] The Tsetlin Machine (TM) is used as the core classifier to classify the one-dimensional Boolean feature vector X to obtain the category of the UAV radio frequency signal. The calculation process of the TM algorithm for classification is as follows: each class i has a series of clauses. As voters, those who cast more votes will be the final system output. The specific steps are as follows:

[0097] Step S7.1: The input data is a one-dimensional Boolean feature vector. To simultaneously capture both the "presence" and "absence" states of features, the TM algorithm first converts the one-dimensional Boolean feature vector X and its logical NOT feature... Perform splicing to generate a length of The set of literals :

[0098]

[0099] Step S7.2: Each clause is for each element in the literal set. Each was assigned a dedicated Tsetlin automaton. , indicating the first The class of Clauses as elements The assigned TA clause. The output is a logical AND operation consisting of the literals corresponding to all the automata that decide to perform the "include" action. The calculation yields:

[0100]

[0101] Step S7.3: The TM model assigns the same number of logistic clauses to each category, each with opposite polarities:

[0102] Positive polarity clause : Vote in favor of this category upon activation (score) This clause indicates that the input data belongs to this category.

[0103] Negative polarity clause : Vote against this category when activating (score) This indicates that the clause determines that the input data does not belong to this category.

[0104] When completing a reasoning operation, the category Total score Subtract the sum of the outputs of all negative clauses from the sum of the outputs of all positive clauses:

[0105]

[0106] in, The total number of clauses assigned to a single category.

[0107] Step S7.4: The system calculates the scores for all categories in parallel, and then... The function identifies the category with the highest score as the final recognition result:

[0108]

[0109] Example Simulation

[0110] In this embodiment, the present invention uses ten types of UAV radio frequency signal datasets for training. The selected UAV models are shown in Table 1.

[0111]

[0112] Table 1

[0113] Figure 6 Experimental results show that after 50 iterations, the model's classification accuracy on the training set can be stabilized at 95%, and the classification accuracy on the validation set can be stabilized at 91%.

[0114] Figure 7 The confusion matrix model of the 1D-TM model used in this invention is demonstrated. The parameter combination of M=1500 clauses and 50 training epochs was determined. The final model achieved an accuracy of 90.88% on the test set. Analysis from the perspectives of precision, recall, and F1-score shows that the model performs well in recognizing most drone models, but performs poorly in recognizing a few models, with T0010 achieving a precision of only 0.7436 and T0011 achieving a recall of only 0.7111.

[0115] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for identifying radio frequency signals of unmanned aerial vehicles (UAVs) based on the Tsetlin Machine, characterized in that, Includes the following steps: (1) Radio frequency signal preprocessing to generate a two-dimensional time-frequency diagram; First, the radio frequency I / Q signals of the drone are acquired and sampled to obtain the radio frequency sampled signal. And generate a spectrum using short-time Fourier transform. : ； in, It is a real-valued symmetric window function. For the sliding step index on the time axis, Angular frequency; Then, by calculating the spectrum The square of the modulus, i.e. the power spectral density, yields the two-dimensional time-frequency diagram of the UAV's radio frequency signal. : ； (2) Generate the initial binary logic matrix; First, from the two-dimensional time-frequency diagram Random selection Pixels, and calculate The intensity values ​​of the nth pixel are then sorted from low to high, and the nth pixel is selected. The intensity value of each pixel is used as the threshold. , two-dimensional time-frequency graph Binarization yields the initial binary logical matrix. : ； (3) Noise reduction is performed using mathematical morphology filtering; For the initial binary logic matrix Perform mathematical morphological filtering to compress it to Standardized logical matrix ,in, Indicates row and column index; (4) Perform maximum value projection; Standardized logic matrix Transform into time-domain and frequency-domain sequences; 4.1) Time-domain projection to obtain the time-domain sequence. : ； 4.2) Frequency domain projection to obtain the frequency domain sequence. : ； (5) First-order alignment to eliminate time offset; For time-domain sequences ,from Start towards Perform a search at the index of the first logical "1" occurrence, and denote it as... And logically shift the vector to the left. The bits are used to obtain the aligned time-domain vector. : ； (6) Add statistical feature information; 6.1) Extracting the frequency domain sequence The end Invalidate all zeros to obtain the time-domain vector. ; 6.2) Constructing using expert prior statistical features The bit sequence E, where the expert prior statistical features are the number of frequency hopping blocks in the two-dimensional time-frequency diagram, whether there is a large-area image transmission signal in the energy density, and spatial distribution information; 6.3) Generate a one-dimensional Boolean eigenvector of fixed length. : ； (7) Construct a lightweight pure logic recognition model; The Tsetlin Machine is used as the core classifier to classify the one-dimensional Boolean feature vector X to obtain the category of the UAV radio frequency signal.

2. The UAV radio frequency signal identification method based on Tsetlin Machine according to claim 1, characterized in that, Time-domain vector A time-domain vector with a length of 59. The value is 64, thus forming a one-dimensional Boolean eigenvector. It has 128 bits; among which, the expert prior statistical features correspond to bits 124-128 of the one-dimensional Boolean feature vector X, and their specific meanings are as follows: 124 bits are frequency hopping active bits: the number of independent frequency hopping blocks with an area of ​​<50 pixels is counted by analyzing connected components to characterize the frequency hopping density; 125 bits are energy density bits: used to determine whether there is a giant connected region with an area greater than 60 pixels, serving as the physical basis for distinguishing broadband OFDM image transmission signals from narrowband FHSS control signals; Bits 126–128 are spatial preference bits: to determine whether the frequency hopping point is concentrated in the upper or lower half of the observed frequency band, in order to capture the frequency hopping strategy preferences of different manufacturers.

Images