A deep learning-based wide-frequency unmanned aerial vehicle spectrum detection and rapid identification method

Abstract

The present application belongs to the technical field of radio spectrum monitoring, and provides a wide-frequency unmanned aerial vehicle spectrum detection and rapid identification method based on deep learning, which comprises the following steps: collecting wide-frequency radio signals to form a composite signal sequence and converting the composite signal sequence into a time-frequency feature map, performing cepstrum transformation on the original signal to extract a cepstrum time delay peak, constructing an environment compensation parameter vector, identifying frequency selective fading notches in the time-frequency feature map, dividing to form a spectrum fragment set, constructing a fragment association model in combination with a frequency domain spacing and an inverse frequency position matching rule, and completing physical constraints, inputting the spectrum fragment spatial features and the environment compensation parameter vector into the model, calculating a coherence weight through a deep fragment association model based on a Transform architecture, completing logical reorganization of the spectrum fragments in a vector space and generating a logical reorganization vector, inputting the vector into a double-branch parallel classification network, and realizing accurate identification of unmanned aerial vehicle target models and communication protocols.

Description

Technical Field

[0001] This invention belongs to the field of radio spectrum monitoring technology, specifically a method for broadband UAV spectrum detection and rapid identification based on deep learning. Background Technology

[0002] With the rapid development of drone technology, spectrum detection and identification technology for low-altitude, slow-moving, and small targets has become the core of low-altitude security defense. Traditional identification methods are gradually shifting from manual feature extraction to methods based on deep learning models such as convolutional neural networks (CNNs). By collecting time-frequency maps or power spectra of drone radiation signals, deep learning models can automatically learn the texture and envelope features of the signals, thereby achieving high-precision identification of drone models and communication protocols.

[0003] However, in practical applications, especially in environments with strong metal reflection (such as large aircraft hangars, automated metal warehouses, ship cabins, or industrial parks with a large number of metal structures), existing detection methods face serious performance degradation problems.

[0004] First, broadband UAV image transmission signals (such as 20MHz / 40MHz bandwidth signals using OFDM modulation) experience severe frequency-selective fading in strong metallic environments. Because metal surfaces have extremely high reflectivity to electromagnetic waves and minimal absorption loss, the receiver receives a large number of reflected signals arriving via different paths with significant time delays. When the phase difference between these multipath signals approaches 180 degrees, severe destructive interference occurs, forming notches deeper than 20dB within the spectral envelope of the broadband signal.

[0005] Secondly, this physical-level destructive interference causes spectral holes in the originally continuous and flat broadband spectrum, making the complete signal envelope appear as multiple physically broken fragments in the frequency domain.

[0006] Existing deep learning-based recognition models, during design and training, typically assume that the signal is in an open or weak multipath environment, making them highly dependent on the global connectivity and texture integrity of the signal's spectral features. In the aforementioned environment with strong metallic reflection, the features extracted by the convolutional kernels of CNN models are no longer the overall structure of a broadband communication protocol, but are mistakenly identified as a series of discrete, randomly distributed narrowband bursts or industrial noise. This topological distortion of features caused by environmental physical characteristics makes deep learning recognition models highly susceptible to misclassifying normal UAV signals as discontinuous interference signals or environmental noise, leading to a significant increase in the false alarm rate and missed alarm rate, severely limiting the reliability of UAV detection systems in complex industrial and military scenarios.

[0007] Therefore, how to solve the spectral fragmentation caused by frequency-selective fading of broadband signals in environments with strong metal reflection, and how to achieve rapid and accurate identification of distorted signals, is a technical problem that urgently needs to be solved in the field of UAV spectrum monitoring.

[0008] Therefore, this invention provides a method for wideband UAV spectrum detection and rapid identification based on deep learning. Summary of the Invention

[0009] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0010] The technical solution adopted by this invention to solve its technical problem is:

[0011] In one aspect, this invention provides a method for broadband unmanned aerial vehicle (UAV) spectrum detection and rapid identification based on deep learning, comprising:

[0012] Step 1: Collect broadband radio signals in the environment under test, form a composite signal sequence, and convert the composite signal sequence into a time-frequency feature map;

[0013] Step 2: Perform cepstral transformation on the original broadband radio signal to obtain a real cepstral sequence, and extract the cepstral delay peak that reflects the characteristics of the metal reflection environment as an environmental compensation parameter vector;

[0014] Step 3: Identify frequency-selective fading notches in the time-frequency feature map, divide the continuous time-frequency feature map into multiple spectral fragments with independent energy envelopes according to the frequency domain position of the notches, form a spectral fragment set, construct a fragment correlation model and complete physical constraints;

[0015] Step 4: Input the spatial characteristics of the spectrum fragments and the environmental compensation parameter vector into the fragment association model, calculate the coherence weights between different spectrum fragments, complete the logical recombination of spectrum fragments in the vector space, and form a logical recombination vector.

[0016] Step 5: Input the logically recombined vector into the classification network, and output the category and protocol recognition results of the UAV target.

[0017] Preferably, the specific process for forming the composite signal sequence is as follows:

[0018] Using a wideband omnidirectional antenna, wideband radio signals in the test environment are sensed in real time. The sensed wideband radio signals are preprocessed by a limiter and a bandpass filter. The preprocessed wideband radio signals are then fed into a low-noise amplifier (LNA) for gain compensation. A high-coherence local oscillator signal generated by a local oscillator is used to mix the wideband radio signals, which are then down-converted to intermediate frequency (IF) or zero IF. An analog-to-digital converter (ADC) is used to perform quadrature sampling on the down-converted radio signals. The sampling frequency must meet the Nyquist criterion and be more than twice the communication bandwidth of the UAV to obtain a composite signal sequence containing the in-phase component I and the quadrature component Q with phase information.

[0019] Preferably, the specific process of converting the composite signal sequence into a time-frequency feature map is as follows:

[0020] The composite signal sequence is divided into consecutive time frames, and a preset overlap rate is set between adjacent frames. The overlap rate ranges from 25% to 50%. A non-rectangular window function is applied to each frame of the composite signal sequence. The non-rectangular window function is preferably a Hamming window or a Blackman window.

[0021] An N-point Fast Fourier Transform is performed on each frame of the composite signal sequence after applying a non-rectangular window function to convert the time-domain complex signal into frequency-domain complex coefficients. The value of N is determined according to the bandwidth and frequency resolution requirements of the broadband radio signal.

[0022] The instantaneous power spectrum is obtained by calculating the complex coefficients in the frequency domain and taking the square of their modulus. Logarithmic compression is then applied to the instantaneous power spectrum. , The preset minimum smoothing factor, This is the logarithmic power spectrum of a single frame. This refers to common logarithmic operations with base 10. After the Fast Fourier Transform, the squared modulus of the complex coefficients at each frequency point is taken.

[0023] The logarithmic power spectra of a single frame are stacked vertically in chronological order to construct a two-dimensional time-frequency matrix. The two-dimensional time-frequency matrix is ​​then adjusted to a preset pixel size using a bilinear interpolation algorithm and normalized to the global mean and standard deviation to generate a time-frequency feature map.

[0024] Preferably, the specific process for obtaining the real cepstral sequence is as follows:

[0025] Obtain the frequency domain complex coefficients, calculate the power spectrum, take the natural logarithm of the power spectrum, and decouple the product relationship into a linear superposition of the logarithmic spectrum of the UAV's transmitted signal and the logarithmic spectrum of the channel frequency response: , To receive the signal power spectrum, The original transmitted signal power spectrum, The channel frequency response includes strong metallic multipath effects;

[0026] For the decoupled logarithmic power spectrum sequence Perform an inverse fast Fourier transform (IFFT) to map the sequence from the frequency domain to the inverse frequency domain, obtaining the real cepstrum sequence.

[0027] Preferably, the specific process of extracting the cepstral delay peak reflecting the characteristics of the metal reflection environment as an environmental compensation parameter vector is as follows:

[0028] By applying a preset cepstral high-pass window function to the real cepstral sequence, the high cepstral component sequence that purely reflects the physical structure of spatial multipaths is separated.

[0029] Local maxima search is performed in the separated high cepstral component sequence. A dynamic threshold is set, and cepstral delay peaks exceeding the dynamic threshold are extracted. The cepstral delay peak position and its corresponding amplitude are recorded.

[0030] The cepstral delay peaks exceeding the dynamic threshold are sorted from largest to smallest amplitude, and the top K delay peaks are extracted as environmental compensation parameters. K is a preset positive integer with a value range of 3 to 10. An environmental compensation parameter vector reflecting the current strong metal reflection spatial topology is constructed.

[0031] Preferably, the specific process for identifying frequency-selective fading notch waves in the time-frequency feature map is as follows:

[0032] For the time-frequency feature map, energy accumulation projection is performed along the time axis to obtain a one-dimensional average power spectrum curve. The one-dimensional average power spectrum curve is smoothed using a moving average filter. The abrupt change point is detected using a second derivative operator or an adaptive threshold segmentation algorithm.

[0033] Set the fading depth threshold and notch bandwidth threshold;

[0034] When the energy of the one-dimensional average power spectrum curve in a specific frequency band is less than the fading depth threshold and the duration is greater than or equal to the notch bandwidth threshold, it is determined to be a frequency-selective fading notch region.

[0035] Preferably, the specific process of constructing the fragment association model and completing the physical constraints is as follows:

[0036] Based on the location of the fading notch region, the notch region is marked as 0 and the normal spectrum region is marked as 1, and a binary spectrum mask matrix is ​​constructed. In the binary spectrum mask matrix, the energy concentration region of the non-notch region is marked as the effective signal region, and the notch region is marked as the feature missing region.

[0037] The time-frequency feature map is logically divided into multiple non-adjacent sub-matrix blocks in the frequency domain using a binarized spectral mask matrix.

[0038] Each sub-matrix block is independently encapsulated to form a set of spectral fragments. For each spectral fragment, the relative geometric coordinates are extracted.

[0039] Based on relative geometric coordinates, the frequency domain spacing between adjacent spectral fragments is calculated. With reciprocal frequency position satisfy When determining the broken parts belonging to the same physical signal, a topological association is established, the fragment association model is constructed, and physical constraints are built.

[0040] Preferably, the specific process for calculating the coherence weights between different spectral fragments is as follows:

[0041] Each spectral fragment is mapped to a feature vector using a lightweight convolutional network or fully connected layer;

[0042] Using the fragment correlation model as a physical constraint, the topological correlation logic of the determined spectral fragments, as well as the frequency domain spacing and environmental compensation parameter vectors, are extracted. Through a nonlinear mapping network, the fused prior constraints are transformed into environmental feature bias terms. , reversing the frequency position Converting to frequency domain interferometric periods, a deep fragment correlation model based on the Transformer architecture is constructed, which extracts the feature vector sequences of all spectral fragments. As a query vector, key vector, value vector, and environmental feature bias term Input a deep fragment association model, and precisely adjust the association weights between fragments using a coherence evaluation function: , For coherence evaluation function, Let be the frequency domain distance between spectral fragment i and spectral fragment j. As a regulating factor;

[0043] If the frequency domain spacing between two spectral fragments meets the physical constraints of the fragment correlation model, the coherence evaluation function outputs a positive gain, thus completing the accurate calculation of the coherence weight matrix.

[0044] Preferably, the specific process of forming the logical recombination vector is as follows:

[0045] The feature vectors of each spectral fragment are weighted and summed using a coherent weight matrix. Under the physical constraints of the fragment correlation model, the discrete spectral fragment features are aggregated into a global feature vector with global envelope semantics. The aggregated global feature vector is then residually connected with the global statistical features of the original time-frequency feature map to output a logical recombination vector.

[0046] Preferably, the specific process for outputting the category and protocol identification results of the UAV target is as follows:

[0047] The logical recombination vector is input into the feedforward layer of the classification network, which consists of a multilayer perceptron (MLP). The classification network employs a dual-branch parallel structure at its end, independently distinguishing between the target model and the communication protocol.

[0048] Classify the individual characteristics of drone hardware radiation sources to distinguish specific models from different manufacturers;

[0049] Classify the modulation methods, frequency hopping patterns, and frame structure characteristics of broadband radio signals to identify protocols such as OcuSync, Lightbridge, Wi-Fi, or custom spread spectrum protocols;

[0050] The output vectors of each classification branch are mapped to a score vector of the same length as the predefined number of categories through a fully connected layer; the softmax function is then used to transform the score vectors into a probability distribution.

[0051] , The original score for the k-th category. This represents the total number of categories;

[0052] Set a global recognition confidence threshold. If the highest probability value is greater than the global recognition confidence threshold, output the corresponding recognition result.

[0053] On the other hand, the present invention provides a deep learning-based broadband unmanned aerial vehicle (UAV) spectrum detection and rapid identification system, comprising:

[0054] Signal acquisition and time-frequency feature construction module: Acquires broadband radio signals under the test environment, forms a composite signal sequence, and converts the composite signal sequence into a time-frequency feature map;

[0055] Environmental compensation parameter extraction module: Performs cepstral transformation on the original broadband radio signal to obtain a real cepstral sequence, and extracts the cepstral delay peak reflecting the characteristics of the metal reflection environment as an environmental compensation parameter vector;

[0056] Spectrum fragment segmentation and correlation model construction module: Identifies frequency-selective fading notches in the time-frequency feature map, divides the continuous time-frequency feature map into multiple spectrum fragments with independent energy envelopes according to the frequency domain position of the notches, forms a spectrum fragment set, constructs a fragment correlation model and completes physical constraints;

[0057] Spectrum fragment spatial logical reorganization module: The spatial characteristics of spectrum fragments and environmental compensation parameter vectors are input into the fragment association model to calculate the coherence weights between different spectrum fragments, and complete the logical reorganization of spectrum fragments in the vector space to form a logical reorganization vector.

[0058] Target result recognition module: Input the logically recombined vector into the classification network and output the category and protocol recognition results of the UAV target.

[0059] The beneficial effects of this invention are as follows:

[0060] 1. By integrating environmental compensation parameters and physical prior constraints, the interference of fading notch waves on spectral characteristics is effectively suppressed, solving the identification bias problem caused by signal fragmentation in strong reflection environments. The logical reassembly strategy based on the spectral fragment correlation model significantly improves the system's stability in complex electromagnetic environments and reduces the risk of false alarms and missed detections. The adoption of lightweight feature extraction and weighted correlation strategies reduces computational overhead while ensuring identification accuracy, making it suitable for real-time monitoring scenarios.

[0061] 2. Designed for applications with dense metal concentrations and significant multipath effects, it can be widely used in low-altitude security defense, electromagnetic situational awareness, and other fields, demonstrating strong practicality. Through cross-fragment logic stitching, it effectively recovers the global envelope and correlation characteristics of signals damaged by fading, improving the model's ability to represent discontinuous spectra. Attached Figure Description

[0062] The invention will now be further described with reference to the accompanying drawings.

[0063] Figure 1 This is a flowchart illustrating the steps of a broadband UAV spectrum detection and rapid identification method based on deep learning according to the present invention.

[0064] Figure 2 This is a system module diagram of a broadband UAV spectrum detection and rapid identification system based on deep learning, according to the present invention. Detailed Implementation

[0065] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0066] Example 1

[0067] like Figure 1 As shown in the embodiment of the present invention, a method for wideband UAV spectrum detection and rapid identification based on deep learning includes:

[0068] Step 1: Collect broadband radio signals in the environment under test, form a composite signal sequence, and convert the composite signal sequence into a time-frequency feature map;

[0069] In step one, the specific process of acquiring broadband radio signals in the environment under test is as follows:

[0070] Using a wideband omnidirectional antenna, broadband radio signals in the environment under test are sensed in real time.

[0071] The wideband omnidirectional antenna covers a frequency range that includes at least 2.4 GHz to 2.485 GHz and 5.725 GHz to 5.85 GHz, commonly used image transmission frequency bands for UAVs.

[0072] In step one, the specific process of forming the composite signal sequence is as follows:

[0073] The sensed broadband radio signal is preprocessed by a limiter and a bandpass filter to filter out strong out-of-band interference and prevent multipath gain from causing receiver front-end saturation in environments with strong metal reflection.

[0074] The pre-processed broadband radio signal enters a low-noise amplifier (LNA) for gain compensation to improve signal sensitivity at frequency-selective fading notch.

[0075] The high-coherence local oscillator signal generated by the local oscillator is mixed with the broadband radio signal to downconvert the radio signal to an intermediate frequency signal or a zero intermediate frequency signal.

[0076] A high-bit-width, high-sampling-rate analog-to-digital converter (ADC) is used to perform quadrature sampling on the down-converted radio signal. The sampling frequency must meet the Nyquist criterion and be more than twice the communication bandwidth of the UAV, so as to obtain a composite signal sequence containing in-phase (I) and quadrature (Q) components with phase information.

[0077] In step one, the specific process of converting the composite signal sequence into a time-frequency feature map is as follows:

[0078] The acquired composite signal sequence is divided into continuous time frames. In order to ensure the complete capture of UAV frequency hopping signals and instantaneous burst features, a preset overlap rate is set between adjacent frames. The overlap rate ranges from 25% to 50%.

[0079] A non-rectangular window function is applied to each frame of the composite signal sequence. The non-rectangular window function is preferably a Hamming window or a Blackman window. The windowing process suppresses spectral leakage caused by signal truncation, thereby ensuring that the residual signal envelope features can still be clearly distinguished at the edge of the Notch generated by strong metal multipath interference.

[0080] An N-point Fast Fourier Transform is performed on each frame of the composite signal sequence after applying a non-rectangular window function to convert the time-domain complex signal into frequency-domain complex coefficients; wherein, the value of N is determined according to the bandwidth and frequency resolution requirements of the broadband radio signal, preferably 1024, 2048 or 4096;

[0081] The instantaneous power spectrum is obtained by squared the modulus of the complex coefficients at each frequency point (corresponding to the frequency component) after the Fast Fourier Transform. Considering the drastic fluctuations in the dynamic range of signals under strong metal reflection environments, logarithmic compression (dB mapping) is applied to the instantaneous power spectrum, as shown in the following formula:

[0082] ;

[0083] in:

[0084] The preset minimal smoothing factor is used to prevent numerical calculation overflow caused by low energy at the notch. At the same time, the weak signal texture in the notch region is enhanced by logarithmic compression, so that it can be perceived by deep learning networks in terms of visual features. This is the logarithmic power spectrum of a single frame, in dB. 10 represents the logarithmic conversion factor. This refers to common logarithmic operations with base 10. After the Fast Fourier Transform, the squared modulus of the complex coefficients at each frequency point is taken.

[0085] The logarithmic power spectra of multiple consecutive frames are stacked vertically in chronological order to construct a two-dimensional time-frequency matrix. The two-dimensional time-frequency matrix is ​​adjusted to a preset pixel size (such as 224×224 or 256×256) using a bilinear interpolation algorithm, and global mean standard deviation normalization is performed to finally generate a high-resolution time-frequency feature map for input to a deep learning model.

[0086] Step 2: Perform cepstral transformation on the original broadband radio signal to obtain a real cepstral sequence, and extract the cepstral delay peak that reflects the characteristics of the metal reflection environment as an environmental compensation parameter vector;

[0087] In step two, the specific process of performing cepstral transform on the original broadband radio signal to obtain the real cepstral sequence is as follows:

[0088] Obtain the frequency domain complex coefficients of the Fast Fourier Transform (FFT) output in step one, and calculate its power spectrum;

[0089] Because in a strong metal reflection environment, the power spectrum of the received broadband radio signal is the product of the power spectrum of the UAV's transmitted signal and the frequency response of the multipath channel;

[0090] Taking the natural logarithm of the power spectrum, the product relationship is decoupled into a linear superposition of the logarithmic spectrum of the UAV's transmitted signal and the logarithmic spectrum of the channel frequency response, as shown in the following formula:

[0091] ;

[0092] in, To receive the signal power spectrum, The original transmitted signal power spectrum, The channel frequency response includes strong metallic multipath effects;

[0093] For the decoupled logarithmic power spectrum sequence Performing an inverse fast Fourier transform (IFFT) maps the sequence from the frequency domain to the quefrency domain, yielding a real cepstrum sequence. :

[0094] ;

[0095] In the reciprocal frequency domain, the periodic notch characteristics of the broadband radio signal under test caused by multipath interference in the frequency domain will be transformed into discrete impact peaks on the cepstral axis; which can be effectively separated from the stationary spectral characteristics of the broadband radio signal originally transmitted by the UAV.

[0096] In step two, the specific process of extracting the cepstral delay peak value reflecting the characteristics of the metal reflection environment as the environmental compensation parameter vector is as follows:

[0097] Since the envelope characteristics of the broadband signal corresponding to the original transmission signal of the UAV are mainly concentrated in the low cefrency band of the cepstral spectrum, while the multipath delay characteristics caused by reflection from the metallic environment are distributed in the mid-to-high cefrency band.

[0098] For the real cepstral sequence By applying a preset cepstral high-pass window function, the low-frequency cepstral components that represent the macroscopic envelope of the signal itself are filtered out, and the high-frequency cepstral component sequence that purely reflects the physical structure of the spatial multipath is separated out. ;

[0099] In the separated high-frequency inverse component sequence Local maxima search is performed, and a dynamic threshold is set (e.g., 1.5 to 3 times the mean of the current high-frequency cepstral component sequence is used as the amplitude judgment threshold) to eliminate background noise interference; cepstral delay peaks exceeding the dynamic threshold are extracted, and the cepstral delay peak position is recorded. and its corresponding amplitude ;

[0100] Wherein, the inverted frequency position It has a time dimension and directly characterizes the physical multipath delay of the i-th metal reflection path relative to the direct path; the amplitude This characterizes the energy intensity of the reflection path;

[0101] The cepstral delay peaks exceeding the dynamic threshold are sorted by amplitude from largest to smallest, and the top K significant cepstral delay peaks are extracted as environmental compensation parameters, where K is a preset positive integer, ranging from 3 to 10, preferably 5, to construct an environmental compensation parameter vector reflecting the current strong metal reflection spatial topology. :

[0102] ;

[0103] The environmental compensation parameter vector The physical environment constraints will be used as prior knowledge and input into the subsequent deep learning model to guide the model to perform cross-frequency band logical stitching of spectrum fragments broken by notch filtering.

[0104] Step 3: Identify frequency-selective fading notches in the time-frequency feature map, divide the continuous time-frequency feature map into multiple spectral fragments with independent energy envelopes according to the frequency domain position of the notches, form a spectral fragment set, construct a fragment correlation model and complete physical constraints;

[0105] In step three, the specific process of identifying frequency-selective fading notch waves in the time-frequency feature map is as follows:

[0106] For the time-frequency feature map (waterfall plot) generated in step one, energy accumulation projection is performed along the time axis (that is, the energy values ​​of the same frequency point at different time frames are arithmetically accumulated) to obtain a one-dimensional average power spectrum curve reflecting the energy distribution of broadband radio signals within the current observation period. The curve is smoothed using a moving average filter to eliminate instantaneous noise fluctuations and extract the power spectral envelope that reflects the macroscopic structure of the signal.

[0107] The abrupt change point of the one-dimensional average power spectrum curve is detected using the second derivative operator or an adaptive threshold segmentation algorithm;

[0108] Set fading depth threshold (e.g., a drop of more than 15 dB relative to peak energy) and notch bandwidth threshold ;

[0109] When the energy drop of the one-dimensional average power spectrum curve in a specific frequency band exceeds (or is less than) the fading depth threshold And the sustained width reaches (greater than or equal to) the notch bandwidth threshold. At that time, it was determined to be a frequency-selective fading notch region caused by strong metal multipath interference; the starting frequency of each notch region was recorded. and cutoff frequency ;

[0110] In step three, the continuous time-frequency feature map is divided into multiple spectral fragments with independent energy envelopes according to the frequency domain location of the notch, forming a spectral fragment set. The specific process of constructing the fragment correlation model and completing the physical constraints is as follows:

[0111] Based on the identified frequency-selective fading notch region locations, the notch region is marked as 0 and the normal spectrum region is marked as 1, and a binarized spectrum mask matrix is ​​constructed.

[0112] In the binarized spectral mask matrix, the energy concentration region of the non-notch region is marked as the effective signal region, and the notch region is marked as the feature missing region.

[0113] The original time-frequency feature map is logically divided using the binarized spectrum mask matrix, thereby decomposing the original continuous but cut-off broadband radio signal physical strip structure into multiple non-adjacent sub-matrix blocks in the frequency domain.

[0114] Each sub-matrix block is independently encapsulated to form a set of spectral fragments. For each spectrum fragment Extract its local feature parameters, including: the center frequency of the fragment. relative bandwidth Average energy intensity within the fragment And the relative geometric coordinates of the fragment in the time-frequency plot;

[0115] Based on the relative geometric coordinates of each spectral fragment in the time-frequency feature map, the frequency domain spacing between adjacent spectral fragments is calculated. (For example, take the absolute value of the difference between the corresponding coordinate points of the two fragments on the frequency axis); compare the frequency domain spacing with the inverse frequency position in the environmental compensation parameter vector extracted in step two. Pre-comparison is performed to initially establish the topological correlation logic between different fragments, that is, when the frequency domain spacing... With reciprocal frequency position satisfy When the two fragments are determined to be the broken parts of the same physical signal, a topological association is established, the fragment association model is constructed, and physical constraints are built.

[0116] in, The frequency domain spacing is the i-th frequency domain spacing;

[0117] The relative geometric coordinates in the time-frequency feature map and the spectral fragment set will be used as the structured input for the subsequent cross-hole feature stitching step to solve the recognition failure problem caused by feature breakage in deep learning models.

[0118] Step 4: Input the spatial characteristics of the spectrum fragments and the environmental compensation parameter vector into the fragment association model, calculate the coherence weights between different spectrum fragments, complete the logical recombination of spectrum fragments in the vector space, and form a logical recombination vector.

[0119] In step four, the spatial characteristics of the spectral fragments and the environmental compensation parameter vector are input into the fragment correlation model. The specific process for calculating the coherence weights between different spectral fragments is as follows:

[0120] Using lightweight convolutional networks or fully connected layers, each spectral fragment extracted in step three is processed... Mapped to feature vectors ;

[0121] The feature vector The local texture, energy envelope, and frequency domain relative position information of the fragments are encoded. At the same time, positional encoding is introduced to preserve the absolute frequency order of each fragment in the original wideband bandwidth, forming a standardized spectral fragment feature vector sequence, which serves as the basic input for deep correlation calculation.

[0122] Using the fragment association model constructed in step three as the core physical prior constraint, the topological association logic of the spectral fragments and the frequency domain spacing matching rules determined in the model are extracted. (and the environmental compensation parameter vector output in step two) (Including inverted frequency position) and amplitude Then, a nonlinear mapping network is used to transform the fused prior constraints into environmental feature bias terms. Based on the principle of multipath interference, the inverted frequency position is... Converting to frequency domain interference periods further strengthens the topological correlation logic constraints between spectral fragments of the same radiation source;

[0123] A deep fragment association model based on the Transformer architecture is constructed, with the multi-head self-attention operator at its core; the feature vector sequences of all spectral fragments are then used. This serves as the input model for the query vector, key vector, and value vector; it is used in calculating the attention score matrix. At that time, the environmental feature bias term that incorporates the fragment association model from step three is then used. Direct injection, with precise adjustment of the correlation weights between fragments through a coherence evaluation function, as shown in the following formula:

[0124] ;

[0125] in, For coherence evaluation function, Let be the frequency domain distance between spectral fragment i and spectral fragment j. As a regulating factor;

[0126] If the frequency domain spacing between two spectral fragments meets the physical constraints of the fragment correlation model, the coherence evaluation function outputs a positive gain, forcibly increasing the correlation weight between fragments that cross the frequency selective fading notch hole, thus completing the accurate calculation of the coherence weight matrix.

[0127] In step four, the logical recombination of spectral fragments in the vector space is completed, and the specific process of forming the logically recombined vector is as follows:

[0128] The calculated coherent weight matrix is ​​used to perform a weighted summation of the feature vectors of each spectrum fragment. Under the physical topological constraints of the fragment association model in step three, the precise logical stitching across the notch in the feature space is realized, and the discrete spectrum fragment features are aggregated into global feature vectors with global envelope semantics. This allows the discontinuous features that were originally cut off by the notch to be reassembled into a connected whole that conforms to the characteristics of the UAV communication protocol in strict accordance with physical laws.

[0129] Simultaneously, the aggregated global feature vector is joined with the global statistical features of the original time-frequency feature map generated in step one to compensate for the background noise features that may be lost during the fragmentation process.

[0130] The final output logical recombination vector integrates the physical prior constraints of step three, the UAV signal modulation attributes, and the environmental perception capability of strong metal multipath reflection. It not only makes full use of the fragment association model constructed in step three, but also achieves refined association through deep model, effectively eliminating the ambiguity of deep learning model in identifying broken signals, and providing a global feature foundation with high discriminative power and strong physical rationality for subsequent classification and recognition.

[0131] Step 5: Input the logically recombined vector into the classification network, and output the category and protocol recognition results of the UAV target.

[0132] In step five, the specific process of inputting the logically recombined vector into the classification network and outputting the category and protocol identification results of the UAV target is as follows:

[0133] Input the logical recombination vector output from step four into the feedforward layer of the classification network;

[0134] The feedforward layer consists of a multilayer perceptron (MLP), which maps the logically recombined vector features to a more discriminative semantic space through linear transformations and nonlinear activation functions (such as ReLU or GELU).

[0135] This process aims to further compress the residual environmental noise generated by strong metal reflections and extract physical layer features that reflect the essence of the UAV communication protocol.

[0136] Parallel processing of multi-task classification heads: The classification network ends with a dual-branch parallel structure, independently distinguishing between the target model and communication protocol.

[0137] Model identification branch: Classifies individual characteristics of drone hardware radiation sources to distinguish specific models from different manufacturers such as DJI and Autel;

[0138] Protocol identification branch: Classifies the modulation methods, frequency hopping patterns, and frame structure characteristics of broadband radio signals to identify protocols such as OcuSync, Lightbridge, Wi-Fi, or custom spread spectrum protocols;

[0139] Through a multi-task learning framework, the two branches share the underlying global feature vector, thereby improving the model's generalization ability in complex electromagnetic environments.

[0140] The output vectors of each classification branch are mapped to a score vector of the same length as the predefined number of categories through a fully connected layer; the softmax function is then used to transform the score vectors into a probability distribution.

[0141] ;

[0142] in, The original score for the k-th category. The total number of categories; this probability value represents the model's confidence that the current fragmented and reassembled signal belongs to a specific UAV target.

[0143] Set global recognition confidence threshold If the highest probability value Greater than the global recognition confidence threshold If, then output the corresponding recognition result; if Below the global recognition confidence threshold Then, a second verification is performed using the environmental compensation parameter vector extracted in step two.

[0144] If environmental parameters indicate that the current multipath fading is extremely severe (the number of notches exceeds the preset value), the system will automatically activate the adaptive weighting mechanism to reduce the dependence on the broken edge features and recalculate the classification probability to prevent severely distorted UAV signals from being misjudged as noise.

[0145] The system ultimately outputs the identified drone target category (e.g., "DJI Mavic 3") and the communication protocol used (e.g., "OcuSync 3.0"). Based on the identification results, the system updates the detection log in real time and displays the identified drone model icon and signal confidence level on the monitoring interface.

[0146] If the identification result belongs to a blacklist target, the alarm command of the defense system is triggered simultaneously, providing accurate target parameters for subsequent electronic interference or physical countermeasures.

[0147] Example 2

[0148] like Figure 2 As shown in the detailed implementation of Embodiment 1, the present invention provides a deep learning-based wideband UAV spectrum detection and rapid identification system, comprising:

[0149] Signal acquisition and time-frequency feature construction module: Acquires broadband radio signals under the test environment, forms a composite signal sequence, and converts the composite signal sequence into a time-frequency feature map;

[0150] Environmental compensation parameter extraction module: Performs cepstral transformation on the original broadband radio signal to obtain a real cepstral sequence, and extracts the cepstral delay peak reflecting the characteristics of the metal reflection environment as an environmental compensation parameter vector;

[0151] Spectrum fragment segmentation and correlation model construction module: Identifies frequency-selective fading notches in the time-frequency feature map, divides the continuous time-frequency feature map into multiple spectrum fragments with independent energy envelopes according to the frequency domain position of the notches, forms a spectrum fragment set, constructs a fragment correlation model and completes physical constraints;

[0152] Spectrum fragment spatial logical reorganization module: The spatial characteristics of spectrum fragments and environmental compensation parameter vectors are input into the fragment association model to calculate the coherence weights between different spectrum fragments, and complete the logical reorganization of spectrum fragments in the vector space to form a logical reorganization vector.

[0153] Target result recognition module: Input the logically recombined vector into the classification network and output the category and protocol recognition results of the UAV target.

[0154] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for broadband unmanned aerial vehicle (UAV) spectrum detection and rapid identification based on deep learning, characterized in that, include: Step 1: Collect broadband radio signals in the environment under test, form a composite signal sequence, and convert the composite signal sequence into a time-frequency feature map; Step 2: Perform cepstral transformation on the original broadband radio signal to obtain a real cepstral sequence, and extract the cepstral delay peak that reflects the characteristics of the metal reflection environment as an environmental compensation parameter vector; Step 3: Identify frequency-selective fading notches in the time-frequency feature map, divide the continuous time-frequency feature map into multiple spectral fragments with independent energy envelopes according to the frequency domain position of the notches, form a spectral fragment set, construct a fragment correlation model and complete physical constraints; Step 4: Input the spatial characteristics of the spectrum fragments and the environmental compensation parameter vector into the fragment association model, calculate the coherence weights between different spectrum fragments, complete the logical recombination of spectrum fragments in the vector space, and form a logical recombination vector. Step 5: Input the logically recombined vector into the classification network, and output the category and protocol recognition results of the UAV target.

2. The method for broadband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process for forming the composite signal sequence is as follows: Using a wideband omnidirectional antenna, wideband radio signals in the test environment are sensed in real time. The sensed wideband radio signals are preprocessed by a limiter and a bandpass filter. The preprocessed wideband radio signals are then fed into a low-noise amplifier (LNA) for gain compensation. A high-coherence local oscillator signal generated by a local oscillator is used to mix the wideband radio signals, which are then down-converted to intermediate frequency (IF) or zero IF. An analog-to-digital converter (ADC) is used to perform quadrature sampling on the down-converted radio signals. The sampling frequency must meet the Nyquist criterion and be more than twice the communication bandwidth of the UAV to obtain a composite signal sequence containing the in-phase component I and the quadrature component Q with phase information.

3. The method for broadband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process of converting the composite signal sequence into a time-frequency feature map is as follows: The composite signal sequence is divided into consecutive time frames, and a preset overlap rate is set between adjacent frames. The overlap rate ranges from 25% to 50%. A non-rectangular window function is applied to each frame of the composite signal sequence. The non-rectangular window function is preferably a Hamming window or a Blackman window. An N-point Fast Fourier Transform is performed on each frame of the composite signal sequence after applying a non-rectangular window function to convert the time-domain complex signal into frequency-domain complex coefficients. The value of N is determined according to the bandwidth and frequency resolution requirements of the broadband radio signal. The instantaneous power spectrum is obtained by calculating the complex coefficients in the frequency domain and taking the square of their modulus. Logarithmic compression is then applied to the instantaneous power spectrum. , The preset minimum smoothing factor, This is the logarithmic power spectrum of a single frame. This refers to common logarithmic operations with base 10. After the Fast Fourier Transform, the complex coefficients in the frequency domain at each frequency point are squared modulo-1. The logarithmic power spectra of a single frame are stacked vertically in chronological order to construct a two-dimensional time-frequency matrix. The two-dimensional time-frequency matrix is ​​then adjusted to a preset pixel size using a bilinear interpolation algorithm and normalized to the global mean and standard deviation to generate a time-frequency feature map.

4. The method for broadband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process for obtaining the real cepstral sequence is as follows: Obtain the frequency domain complex coefficients, calculate the power spectrum, take the natural logarithm of the power spectrum, and decouple the product relationship into a linear superposition of the logarithmic spectrum of the UAV's transmitted signal and the logarithmic spectrum of the channel frequency response: , To receive the signal power spectrum, The original transmitted signal power spectrum, The channel frequency response includes strong metallic multipath effects; For the decoupled logarithmic power spectrum sequence Perform an inverse fast Fourier transform (IFFT) to map the sequence from the frequency domain to the inverse frequency domain, obtaining the real cepstrum sequence.

5. The method for broadband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process of extracting the cepstral delay peak value reflecting the characteristics of the metal reflection environment as an environmental compensation parameter vector is as follows: By applying a preset cepstral high-pass window function to the real cepstral sequence, the high cepstral component sequence that purely reflects the physical structure of spatial multipaths is separated. Local maxima search is performed in the separated high cepstral component sequence. A dynamic threshold is set, and cepstral delay peaks exceeding the dynamic threshold are extracted. The cepstral delay peak position and its corresponding amplitude are recorded. The cepstral delay peaks exceeding the dynamic threshold are sorted from largest to smallest amplitude, and the top K delay peaks are extracted as environmental compensation parameters. K is a preset positive integer with a value range of 3 to 10. An environmental compensation parameter vector reflecting the current strong metal reflection spatial topology is constructed.

6. The method for broadband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process for identifying frequency-selective fading notch waves in the time-frequency feature map is as follows: For the time-frequency feature map, energy accumulation projection is performed along the time axis to obtain a one-dimensional average power spectrum curve. The one-dimensional average power spectrum curve is smoothed using a moving average filter. The abrupt change point is detected using a second derivative operator or an adaptive threshold segmentation algorithm. Set the fading depth threshold and notch bandwidth threshold; When the energy of the one-dimensional average power spectrum curve in a specific frequency band is less than the fading depth threshold and the duration is greater than or equal to the notch bandwidth threshold, it is determined to be a frequency-selective fading notch region.

7. The method for broadband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process of constructing the fragment association model and completing the physical constraints is as follows: Based on the location of the fading notch region, the notch region is marked as 0 and the normal spectrum region is marked as 1, and a binary spectrum mask matrix is ​​constructed. In the binary spectrum mask matrix, the energy concentration region of the non-notch region is marked as the effective signal region, and the notch region is marked as the feature missing region. The time-frequency feature map is logically divided using a binarized spectral mask matrix, decomposing it into multiple non-adjacent sub-matrix blocks in the frequency domain; Each sub-matrix block is independently encapsulated to form a set of spectral fragments. For each spectral fragment, the relative geometric coordinates are extracted. Based on relative geometric coordinates, the frequency domain spacing between adjacent spectral fragments is calculated. With reciprocal frequency position satisfy When determining the broken parts belonging to the same physical signal, a topological association is established, the fragment association model is constructed, and physical constraints are built.

8. The method for broadband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process for calculating the coherence weights between different spectral fragments is as follows: Each spectral fragment is mapped to a feature vector using a lightweight convolutional network or fully connected layer; Using the fragment correlation model as a physical constraint, the topological correlation logic of the determined spectral fragments, as well as the frequency domain spacing and environmental compensation parameter vectors, are extracted. Through a nonlinear mapping network, the fused prior constraints are transformed into environmental feature bias terms. Reverse frequency position Converting to frequency domain interferometric periods, a deep fragment correlation model based on the Transformer architecture is constructed, which extracts the feature vector sequences of all spectral fragments. As a query vector, key vector, value vector, and environmental feature bias term Input a deep fragment association model, and precisely adjust the association weights between fragments using a coherence evaluation function: , For coherence evaluation function, Let be the frequency domain distance between spectral fragment i and spectral fragment j. As a regulating factor; If the frequency domain spacing between two spectral fragments conforms to the physical constraints of the fragment correlation model, the coherence evaluation function outputs a positive gain, thus completing the accurate calculation of the coherence weight matrix.

9. The method for wideband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process for forming the logical recombination vector is as follows: The feature vectors of each spectral fragment are weighted and summed using a coherent weight matrix. Under the physical constraints of the fragment correlation model, the discrete spectral fragment features are aggregated into a global feature vector with global envelope semantics. The aggregated global feature vector is then residually connected with the global statistical features of the original time-frequency feature map to output a logical recombination vector.

10. A method for wideband UAV spectrum detection and rapid identification based on deep learning according to claim 1, characterized in that, The specific process for outputting the category and protocol identification results of the UAV target is as follows: The logical recombination vector is input into the feedforward layer of the classification network, which consists of a multilayer perceptron (MLP). The classification network employs a dual-branch parallel structure at its end, independently distinguishing between the target model and the communication protocol. Classify the individual characteristics of drone hardware radiation sources to distinguish specific models from different manufacturers; Classify the modulation methods, frequency hopping patterns, and frame structure characteristics of broadband radio signals to identify protocols such as OcuSync, Lightbridge, Wi-Fi, or custom spread spectrum protocols; The output vectors of each classification branch are mapped to a score vector of the same length as the predefined number of categories through a fully connected layer; the softmax function is then used to transform the score vectors into a probability distribution. , The original score for the k-th category. Total number of categories; Set a global recognition confidence threshold. If the highest probability value is greater than the global recognition confidence threshold, output the corresponding recognition result.

Images