Identification and analysis method for LoRa unmanned aerial vehicle communication signal

Abstract

The invention relates to the technical field of wireless communication signal big data analysis, and discloses a LoRa unmanned aerial vehicle communication signal-oriented identification and analysis method, which comprises the steps of obtaining radio signal discrete data in a monitoring area, executing short-time Fourier transform to generate a time-frequency two-dimensional spectrogram, and identifying and intercepting a LoRa signal noisy IQ time sequence by using a convolutional neural network; mapping the sequence into a high-dimensional feature tensor, inputting the high-dimensional feature tensor into a slope statistical attention generative adversarial network, calculating a first-order differential tensor by using a generator, carrying out kernel convolution on the first-order differential tensor and preset linearity statistics to generate an attention mask, and reconstructing a de-noised IQ time sequence by using a bidirectional long-short-term memory network after Hadamard product gating; according to the method, the tensor gating operator based on the physical statistical law is embedded in the neural network, so that the common problem that physical semantic drift is easily generated by a generative model at an extremely low signal-to-noise ratio is solved.

Description

Technical Field

[0001] This invention relates to a method for identifying and analyzing communication signals of LoRa drones, belonging to the field of big data analysis technology for wireless communication signals. Background Technology

[0002] Currently, LoRa technology possesses characteristics such as long-distance transmission, low power consumption, and strong anti-interference, making it the mainstream communication solution for telemetry and control links of small UAVs. It adopts linear frequency modulation spread spectrum modulation and uses chirp signals with frequencies that change linearly with time to carry information, maintaining communication connections under low signal-to-noise ratio conditions. Existing monitoring mostly relies on energy detection or cross-correlation matching based on standard preambles. In scenarios with a clean background electromagnetic environment and moderate signal-to-noise ratio, target signal discovery and parameter estimation are completed by detecting signal energy abrupt changes or matching preset standard waveform characteristics. For example, Chinese invention patent CN114397913B discloses a rocket debris search and positioning system and method. The solution deploys a LoRa gateway module and a terminal module on the UAV platform to achieve wireless backhaul and collaborative positioning of infrasound sensor data search network. Although it verifies the application value of LoRa UAV networking, the core logic lies in the scheduling, navigation, and control of system functional modules, while the signal processing layer relies on the conventional demodulation mechanism of standard communication hardware.

[0003] In real-world adversarial scenarios, communication links face complex electromagnetic environments with strong background noise, multipath effects, and co-channel interference. The received signal-to-noise ratio is extremely low or even submerged in the noise floor. Conventional energy detection operators struggle to define the boundary between signal and noise energy, easily leading to false alarms or missed detections. Standard template matching methods suffer from signal transmission distortion or interference masking, resulting in blurred related peaks and making it difficult to obtain accurate spreading factor, bandwidth, and symbol period parameters.

[0004] Therefore, the technical problem to be solved by this invention is to construct a data calculation model based on strong constraints of physical layer signal characteristics, realize high-fidelity reconstruction and accurate parameter analysis of LoRa UAV communication signals under extremely low signal-to-noise ratio and complex interference environment, and avoid feature semantic drift in the data processing process. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for identifying and parsing communication signals for LoRa drones, comprising the following steps: Step S101: Obtain discrete radio signal data within the monitoring area, perform short-time Fourier transform on the discrete radio signal data to generate a time-frequency two-dimensional spectrum, and use a convolutional neural network to identify texture features in the two-dimensional spectrum and extract the noisy IQ time series that is determined to be a LoRa signal. Step S102: The truncated noisy IQ time series is mapped to a high-dimensional feature tensor through a one-dimensional convolutional layer, and the high-dimensional feature tensor is input into the generator in the slope statistical attention generative adversarial network. In step S103, the generator performs a gated computation based on statistical attention on the high-dimensional feature tensor according to the physical statistical laws of the linear frequency modulated signal: the gradient of the high-dimensional feature tensor in the time dimension is calculated to generate a first-order difference tensor; a convolution operation is performed on the first-order difference tensor using a preset linearity statistical kernel to generate a linearity confidence map representing the consistency of the frequency change rate. The linearity statistical kernel is a weight matrix constructed based on the slope distribution statistical features of the noiseless LoRa signal samples; then, the linearity confidence map is mapped to a binary attention mask using an activation function; finally, the Hadamard product of the high-dimensional feature tensor and the attention mask is calculated, and a gated feature tensor with nonlinear noise features filtered out is output. Step S104: Input the gated feature tensor into the bidirectional long short-term memory network to reconstruct the temporal phase dependency of the signal, and map the features back to the dimensional numerical space through the transposed convolutional layer to output the denoised IQ time series. Step S105: Perform fast Fourier transform and autocorrelation analysis on the denoised IQ time series to extract bandwidth and symbol period parameters, and calculate the spreading factor based on the product relationship between bandwidth and symbol period.

[0006] Preferably, the process of constructing the linearity statistical kernel involved in step S103 includes: constructing a large data training set of noiseless LoRa signals containing multiple spreading factors and bandwidth configurations; extracting the local gradient features of each signal sample in the time and frequency domains in the large data training set, and calculating the statistical distribution probability density of the local gradient features; optimizing the weight matrix using an adversarial training mechanism so that the weight matrix produces a high response value to the gradient features that conform to the statistical distribution probability density, and establishing the converged weight matrix as the linearity statistical kernel.

[0007] Preferably, the specific operation of calculating the short-time Fourier transform spectrum in step S101 includes: dividing the discrete data of the radio signal into multiple overlapping time windows; performing a fast Fourier transform on each time window to calculate the square of the amplitude spectrum to obtain the power spectral density; and stitching the power spectral densities of multiple time windows on the time axis to generate a two-dimensional grayscale image reflecting the time-frequency energy distribution of the signal.

[0008] Preferably, the specific operation of identifying and extracting the noisy IQ time series using a convolutional neural network in step S101 includes: inputting a two-dimensional grayscale image into a lightweight convolutional neural network, extracting slope texture features in the image using multi-layer convolutional kernels; outputting the probability value of the LoRa signal existing in the current time window through a fully connected layer; when the probability value exceeds a preset decision threshold, backtracking and extracting the corresponding original noisy IQ time series according to the index position of the time window in the original data stream.

[0009] Preferably, the specific operations for reconstructing the temporal phase dependency using a bidirectional long short-term memory network in step S104 include: inputting the gated feature tensor into the forward and backward layers of the bidirectional long short-term memory network step by step; using the forward layer to capture the evolution trend of the signal phase and using the backward layer to capture the correlation dependency between signal symbols; concatenating the hidden state vectors of the forward and backward layers, and mapping the concatenated vector back to the 1D time domain space through a transposed convolutional layer; restoring the signal amplitude using a denormalization operation to generate a denoised IQ time series.

[0010] Preferably, the specific operations for extracting bandwidth and symbol period parameters in step S105 include: performing a fast Fourier transform on the denoised IQ time series to obtain the spectrum, searching for the maximum and second maximum energy points in the spectrum; determining the upper and lower boundary frequencies of the main lobe of the spectrum based on the energy attenuation threshold, and calculating the difference between the upper and lower boundary frequencies as the bandwidth parameter; performing a sliding window autocorrelation operation on the denoised IQ time series, detecting the position of the first peak in the autocorrelation curve, and establishing the time delay corresponding to the position as the symbol period parameter.

[0011] Preferably, the specific operation of calculating the spreading factor in step S105 includes: multiplying the bandwidth parameter by the symbol period parameter to obtain the number of chips in each symbol period; calculating the base-2 logarithm of the number of chips and rounding the result to obtain the spreading factor.

[0012] Preferably, the method further includes a data-level parameter closed-loop verification step: mapping the calculated center frequency, bandwidth, and spreading factor parameters to register configuration instructions defined by the LoRa communication protocol; reading the status flag data returned by the LoRa communication module after executing the channel activity detection mode; if the status flag data indicates that the preamble match is successful, then the parameters are established as the final parsing result; if the status flag data indicates that the match fails, then the spreading factor is fine-tuned with a preset step size, and a new register configuration instruction is generated to repeat the verification process.

[0013] Preferably, the training process of the slope statistical attention generative adversarial network uses a composite loss function for parameter updates, which is defined as: ,in, The denoised signal output by the generator. For true noise-free signals, This indicates the Fourier transform operation. To counteract the loss, This represents the mean square error term. Represents the spectrum consistency loss term. and To balance the weights, a hyperparameter is used; the energy distribution of the generated signal in the frequency domain is constrained to be consistent with that of the real signal by a spectral consistency loss term.

[0014] Preferably, the method is deployed as an embedded operator in the edge computing unit of the UAV countermeasure system. The edge computing unit generates and adjusts the modulation parameters of the interference signal according to the spreading factor, and performs data blocking operation on the communication link of the target UAV. The discrete data of radio signals in the monitoring area comes from the real-time acquisition stream of the software radio peripheral connected to the edge computing unit.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Introduce a spectral morphology difference calculation term into the optimization objective of the generative adversarial network, establish a time-series phase dependence model in conjunction with a bidirectional long short-term memory network, construct a data reconstruction mechanism strongly constrained by the physical layer signal protocol characteristics, and use the statistical law of the frequency change rate of the linear frequency modulated signal as the boundary of the gradient descent penalty of the neural network. In the adversarial training, the generative model structurally separates the signal frequency jump edge from the background high-frequency random noise. The time-frequency dual constraint strategy enables the computational model to suppress the minimization of mean square error when processing input data with extremely low signal-to-noise ratio or strong burst interference, thereby preventing waveform over-smoothing, removing environmental noise interference, and strictly maintaining the physical consistency of the slope and phase continuity of the output sequence spreading factor, thus preventing semantic drift of data reconstruction features.

[0016] 2. Using deep convolutional neural networks, adversarial training is conducted on massive signal samples with various combinations of bandwidth and spreading factors. The network establishes feature screening logic that does not rely on manual preset fixed thresholds. Through the implicit learning of the statistical differences between the linear frequency modulation structure and the gradient distribution of environmental noise by the convolutional kernel weights, a generalized data cleaning operator that automatically adapts to different modulation rates and frequency band configurations is formed. When facing unknown communication links with non-cooperative targets, there is no need to obtain the signal modulation parameters in advance. The signal components are locked and extracted based on the statistical texture features of the data itself, avoiding the risk of signal truncation caused by parameter mismatch in traditional filters. This establishes the generalization ability of the analysis model in complex and variable electromagnetic environments.

[0017] 3. A high-fidelity time-domain sequence is output through a physical constraint model, providing deterministic spectral structure data input for subsequent frequency domain transformation and periodic analysis. This eliminates spurious peak interference caused by noise in the Fast Fourier Transform and autocorrelation calculation stages. The calculated center frequency, bandwidth, and spreading factor parameters are mapped to the communication hardware channel activity detection logic. The hardware-level correlation matching results are used as physical-level verification of the software algorithm estimation. A closed-loop logic from data generation and parameter inversion to hardware confirmation is established. Deterministic physical detection feedback is used to correct potential biases in the stochastic calculation model, ensuring the uniqueness and accuracy of the target communication link parameter analysis results. Attached Figure Description

[0018] Figure 1 This is a graph showing the time-domain autocorrelation analysis and symbol period detection results of the denoised signal in this invention; Figure 2 This is a diagram showing the spectral structure analysis and modulation parameter estimation of the denoised signal in this invention. Figure 3 This is a comparison diagram of the time-frequency texture features of the LoRa signal in ideal and noisy environments according to the present invention; Figure 4 This is a flowchart of the signal reconstruction and closed-loop analysis based on slope statistical attention in this invention. Figure 5 This is a comparison of the spectra of the original noisy signal and the physically faithful reconstructed signal of this invention. Figure 6 This is a schematic diagram of the hierarchical architecture of the system integrating edge computing and hardware feedback in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0020] This invention provides a method for identifying and parsing LoRa drone communication signals, consisting of four cascaded computational stages. The first stage is a signal initial screening stage based on two-dimensional time-frequency features, utilizing a lightweight convolutional neural network to quickly locate potential LoRa signal segments from massive amounts of radio monitoring data. The second stage is a data reconstruction stage based on a slope-statistical attention generative adversarial network, which uses an embedded physical statistical kernel to perform tensor-level gating cleaning of noisy signals. The third stage is a deterministic parameter inversion stage, calculating the signal modulation parameters through joint analysis of the frequency and time domains. The final stage is a closed-loop verification stage based on hardware feedback, using the physical response of the communication protocol stack to confirm the validity of the parsing results. In the signal initial screening stage, the system uses a software-defined radio peripheral to continuously sample the electromagnetic environment within the monitoring area to obtain... For complex IQ baseband signal streams, given the blurred features of the original time-domain data under low signal-to-noise ratio, the system executes a short-time Fourier transform algorithm to map the one-dimensional time-domain data into a two-dimensional grayscale image with time and frequency. The horizontal axis represents time, the vertical axis represents frequency, and the pixel grayscale value represents power spectral density. This two-dimensional image is input to a pre-trained lightweight convolutional neural network classifier containing multiple convolutional kernels to extract linear frequency modulation texture features that appear as slanted lines in the image. The output layer of the classifier calculates the posterior probability value of the LoRa signal in the current time window through the Softmax function. When this probability value exceeds a preset confidence threshold, the system backtracks and extracts the corresponding original noisy IQ time series based on the index position of the time window, marks it as a candidate data segment, and transmits it to the next stage.

[0021] Traditional denoising algorithms often smooth out signal frequency transition edges to address the prevalent Gaussian white noise and co-channel interference in candidate data segments, leading to errors in spreading factor calculation. To address this engineering challenge, this invention introduces a slope statistical attention generative adversarial network (GAN) to reconstruct the data. The generator in this network employs an encoder-decoder architecture and embeds a slope attention gating unit. The generator takes the input dimension as... The noisy IQ time series is mapped to a dimension of through a one-dimensional convolutional layer. High-dimensional feature tensor The slope attention gating unit performs a physical feature consistency calculation on the feature tensor: calculating the feature tensor. The first-order difference in the time dimension generates a gradient tensor representing the rate of change of frequency. ; Using a preset linearity statistical kernel Perform convolution operation on the gradient tensor, and calculate the linearity statistics kernel. This is the weight matrix, whose numerical distribution is established through supervised learning of the slope statistical distribution of massive noiseless LoRa signal samples. It is used to respond to gradient features that conform to standard linear frequency modulation. The convolution operation outputs a linearity confidence map, which is then processed by the Hard-Sigmoid activation function and transformed into a binary attention mask. The system then combines this attention mask with the original feature tensor. The Hadamard product operation, i.e., element-wise multiplication, is performed to zero out the noise feature channels that do not conform to the statistical law of linear frequency modulation. The feature tensor after gating and filtering is input into a bidirectional long short-term memory network. The forward and backward gating recurrent units are used to capture the evolution trend of the signal phase and restore the phase continuity of the time series. Finally, the feature tensor is mapped back to the 1D time domain space through a transposed convolutional layer to output the denoised IQ time series.

[0022] After acquiring a denoised IQ time series with high physical fidelity, the system enters the parameter inversion stage to analyze the key modulation parameters of the UAV communication link. A Fast Fourier Transform is performed on the denoised sequence to obtain the signal spectrum. The system searches for the energy maximum point in the spectrum and sets an energy attenuation threshold based on this maximum point, for example, at -6dB relative to the peak value, to define the upper and lower boundary frequencies of the main lobe of the spectrum. and Bandwidth parameters The calculation is the difference between the upper and lower boundary frequencies, and the center frequency. The arithmetic mean of the upper and lower boundary frequencies is calculated. To extract the symbol period parameter, the system performs a sliding window autocorrelation operation on the denoised sequence. The autocorrelation function reflects the similarity of the signal itself under different time delays. The system detects the time delay corresponding to the first peak in the autocorrelation curve and establishes it as the symbol period. Based on the spread spectrum modulation principle of the LoRa communication protocol, the spreading factor The logarithmic value representing the number of chips contained in each symbol period is determined by the system according to the formula. Calculate the spreading factor, where This indicates a floor operation. This calculation relies on the high-fidelity restoration of the signal's time-frequency structure during the denoising stage, ensuring that the product of the bandwidth and period parameter strictly falls within the integer neighborhood of the base-2 logarithmic domain. To ensure the absolute reliability of the analytical parameters, the system executes a closed-loop verification process based on hardware feedback. The system will then use the calculated center frequency... ,bandwidth and spreading factor The instructions map to the LoRa RF transceiver's registers and put the RF transceiver into channel activity detection mode. In this mode, the RF hardware uses an internal correlator to cross-correlate the received air signal with a preamble sequence generated based on the configuration parameters. If the hardware status register returns a preamble detection interrupt signal, it indicates that the parsed parameters completely match the target UAV's communication link, and the parsing ends. If the hardware indicates a detection failure, the system will initiate a parameter fine-tuning procedure with a preset step size, such as bandwidth. Spreading factor The parameters are traversed and searched, and the channel activity detection process described above is repeated until the correct communication parameters are locked or the traversal ends.

[0023] Example 1: In applications involving UAV communication monitoring in complex electromagnetic environments, the challenge lies in accurately identifying and analyzing weak LoRa communication signals amidst the superposition of high-intensity background noise and co-channel interference. In practical deployments, monitoring equipment often operates under conditions of extremely low signal-to-noise ratio. Traditional energy detection or template matching methods frequently fail due to their inability to distinguish between noise and signal characteristics, leading to missed detections or misjudgments of the target UAV's communication link. This example demonstrates how the present invention utilizes a deep computing architecture constrained by physical layer statistical features to autonomously complete the entire process from signal discovery to parameter analysis without manual intervention. The technical solution of this invention uses a software-defined radio peripheral to acquire data in real time. Electromagnetic spectrum data within the monitoring area is collected, and the time-domain signal stream is mapped into a time-frequency two-dimensional image using short-time Fourier transform. This step transforms the unique linear frequency modulation physical property of LoRa signals into texture features in the image space. Even when the signal energy is weak, the continuous slope structure is still identifiable in the spectrum. A lightweight convolutional neural network is introduced as the first-level screening mechanism. By utilizing the sensitivity to linear frequency modulation texture features, it quickly locks the time window containing potential signals from massive background noise and extracts the corresponding noisy IQ time series. This detection strategy based on time-frequency texture rather than simple energy avoids the failure problem of the traditional energy threshold method in low signal-to-noise ratio environments.

[0024] As the core value interface, the slope statistical attention generative adversarial network (GAN) is activated immediately after the initial signal screening, undertaking the crucial task of physical semantic reconstruction. Faced with the truncated noisy signal, the network's generator does not employ conventional filtering and smoothing. Instead, it activates the embedded slope attention gating unit and performs gradient-based convolution operations on the feature tensor of the input signal according to a preset linearity statistical kernel. This accurately identifies and retains weak signal components that conform to the standard linear frequency modulation slope law. Simultaneously, it generates a binarized mask to suppress noise feature channels that do not conform to this physical law. This mechanism not only removes random noise but, more importantly, strictly maintains the phase continuity and spreading factor slope characteristics of the signal under extremely low signal-to-noise ratio conditions, preventing physical semantic drift caused by excessive smoothing. Finally, based on the reconstructed high-fidelity signal, the center frequency, bandwidth, and symbol period parameters are accurately extracted through fast Fourier transform and autocorrelation analysis. Hardware-in-the-loop verification confirms that the analysis results are a complete match with the target UAV's communication link, thereby achieving reliable identification and analysis of non-cooperative target signals in complex interference scenarios.

[0025] Example 2: This example constructs a hardware-in-the-loop test platform based on software-defined radio peripherals. The signal transmitter of this platform consists of a programmable signal generator and a commercial LoRa communication module, used to generate a target signal with standard spread spectrum modulation characteristics. The receiver uses a USRPB210 software-defined radio peripheral connected to an omnidirectional antenna, responsible for the real-time acquisition and transmission of broadband spectrum data within the monitoring area. The entire test process is conducted in an unshielded open indoor environment, and continuous wave interference and Gaussian white noise generated by another signal generator are introduced in the same frequency band to simulate the harsh electromagnetic conditions with a signal-to-noise ratio as low as -10dB in actual combat scenarios. The experiment verifies the effectiveness of the initial signal screening stage. During the preset monitoring period, the transmitter sends LoRa data packets containing different spreading factors (SF7 to SF12) and bandwidth configurations (125kHz, 250kHz) at random intervals. The receiving system continuously acquires and processes the spectrum data, and uses short-time Fourier transform to generate time-varying signals. Two-dimensional frequency images were input into a lightweight convolutional neural network for classification. Observational data showed that, under a signal-to-noise ratio of -10dB, although the time-domain waveform was completely submerged in the noise substrate, the linear frequency modulation texture features in the two-dimensional spectrogram still exhibited identifiable energy accumulation trajectories. The classifier successfully locked more than 98.5% of the effective signal windows, and the false alarm rate remained at a low level. This result indicates that the deep learning initial screening mechanism based on time-frequency texture features improves the signal detection capability in low signal-to-noise ratio environments compared to traditional energy detection algorithms. For the truncated noisy signals, this embodiment focuses on evaluating the data reconstruction performance of the slope statistical attention generative adversarial network. The system inputs the truncated noisy IQ sequence into the pre-trained generative adversarial network. The slope attention gating unit inside the network performs gradient-based gating calculations on the feature tensor according to the preset linearity statistical kernel. To intuitively demonstrate the processing effect, the processing data of a typical signal sample is shown in Table 1 below.

[0026] Table 1: Comparison of Key Characteristic Parameters in Signal Processing

[0027] As shown in Table 1, the original noisy signal suffers from severe noise interference, making it impossible to directly extract effective modulation parameters. While traditional filtering methods improve the signal-to-noise ratio (SNR), they result in distortion of the spreading factor slope (deviation reaching 15.3%) and disruption of phase continuity. In contrast, the signal reconstructed by the method of this invention not only improves the SNR to 12.8 dB but also exhibits a spreading factor slope deviation of only 0.8% and extremely low phase continuity error. This demonstrates that the slope attention gating mechanism can accurately preserve the physical layer statistical characteristics of the signal while suppressing noise, effectively preventing physical semantic drift. Finally, based on the reconstructed high-fidelity signal, the system executes parameter inversion and closed-loop verification procedures. For the reconstructed signal samples in Table 1, the Fast Fourier Transform and Autocorrelation Analysis modules calculate a bandwidth of 125.1 kHz and a symbol period of 1.025 ms, respectively. The spreading factor is then calculated according to the formula. The system writes the above parameter configuration into the LoRa RF transceiver and starts the channel activity detection mode. The hardware status register then returns a preamble detection success flag, confirming the consistency between the parsed parameters and the actual transmission parameters (SF7, BW125kHz).

[0028] Example 3: This example combines Figures 1 to 6 This paper describes a method for identifying and parsing communication signals for LoRa drones, such as... Figure 1 As shown in the figure, this curve represents the result of performing a sliding window autocorrelation operation on a denoised IQ time series. The aim is to extract the periodic features of the signal through time-domain analysis. The horizontal axis in the figure is marked as the candidate symbol period. (ms)}, the vertical axis is marked as the average correlation score, the blue continuous curve represents the degree of autocorrelation of the signal under different time delays, and two peak features are marked in the curve. The first peak point is marked with a pink asterisk and the value 2.100ms, and the second peak point is marked with a red circle and the value 6.300ms. It shows the technical process of determining the time delay corresponding to the first peak position in the autocorrelation curve as the symbol period parameter.

[0029] like Figure 2 As shown, the horizontal axis is labeled with frequency (kHz), and the vertical axis with normalized power (dB). The blue solid curve in the figure represents the normalized spectrum, with its main lobe energy clearly visible. The red dashed horizontal line represents the -6dB threshold, used to extract the effective energy range of the main lobe. The two green dashed vertical lines represent estimated bandwidth boundary 1 and estimated bandwidth boundary 2, respectively. The difference between them on the horizontal axis corresponds to the bandwidth parameter. The pink dashed vertical line located between the two green dashed lines represents the estimated center frequency. The complete diagram illustrates the physical basis for determining the spectral main lobe boundary and calculating the bandwidth and center frequency based on the energy attenuation threshold. Figure 3As shown, the figure contains two side-by-side grayscale images to compare the time-frequency two-dimensional features of the signal under ideal and real-world environments. The left image, marked at the top, is a noise-free image that clearly presents a white linear frequency modulation slope texture on a black background. The right image, marked at the top, is a noisy image that shows the original texture features partially submerged in diffuse gray-white random noise spots. This visually reflects the masking effect of environmental noise on the linear texture features of the LoRa signal after the discrete data of the radio signal is generated into a time-frequency two-dimensional spectrum through short-time Fourier transform.

[0030] like Figure 4 As shown, the process begins with the radio signal acquisition and preprocessing stage, which generates a two-dimensional time-frequency spectrogram (STFT) and transmits it to the next stage, the convolutional neural network (CNN) signal recognition and truncation stage, which extracts the noisy IQ time series. This series is then fed into the slope statistical attention generative adversarial network (GAN) reconstruction module. During this process, the system uses a linearity statistical kernel built based on physical statistical laws to perform convolutional gating on the signal, thereby generating a denoised IQ time series and eliminating semantic drift. The processed data then enters the parameter inversion and analysis stage, where the bandwidth, symbol period, and spreading factor are calculated. These parameters are mapped into instructions to drive the hardware closed-loop verification stage, namely the LoRa module channel activity detection. If the status feedback indicates a matching failure, the spreading factor is fine-tuned and re-verified. If the preamble match is successful, the final analysis result is established and the physical fidelity parameters are output. Figure 5 As shown, the horizontal axis represents the frequency index, and the vertical axis represents the power spectral density in decibels (dB). The original signal spectrum, marked by the solid line, exhibits high-frequency and violent sawtooth-like fluctuations, indicating the interference of noise on the signal energy distribution. In contrast, the processed spectrum, marked by the dashed line, shows a relatively smooth and clearly trending envelope shape. The difference in amplitude between the two within the entire frequency index range intuitively reflects the algorithm's suppression of noise components and its reconstruction effect on the effective signal spectrum structure.

[0031] like Figure 6 As shown, the overall system architecture is divided into three logical layers. The data acquisition terminal at the physical layer includes a software radio peripheral and an omnidirectional receiving antenna module. Its main function is to collect environmental signals and output discrete IQ data streams to the subsequent layers. The edge computing analysis unit at the processing layer is equipped with an embedded high-performance processor. This processor integrates a convolutional neural network recognition module, a slope statistical attention reconstruction module, and a parameter inversion calculation module to execute the core algorithm and send register configuration instructions to the execution layer. The hardware verification and countermeasure terminal at the execution layer includes a communication protocol hardware module and an interference signal generator. This terminal performs channel activity detection or interference operations according to the received configuration instructions and feeds back the channel activity status to the edge computing analysis unit in real time, thereby building a complete closed loop of data acquisition, processing and analysis, and hardware verification between the three layers.

[0032] Example 4: Addressing potential algorithmic path issues in the linearity statistical kernel construction process of the slope statistical attention generative adversarial network involved in this invention, this example provides a systematic construction procedure. The procedure defines the initial state of data acquisition and uses MATLAB or Python signal processing toolboxes to generate a standard noiseless chirp signal sample set containing various LoRa spreading factors (SF7 to SF12) and bandwidth configurations (125kHz, 250kHz, 500kHz). The sample set covers uplink and downlink chirps across the entire frequency band, and the sampling rate is set to at least 10 times the signal bandwidth to ensure the integrity of time-frequency characteristics. For each generated time-domain signal sample... The system performs a short-time Fourier transform to map it into a two-dimensional time-frequency matrix. ,in Indicates a time index. This represents the frequency index; based on this, the system performs local gradient feature extraction operations for the matrix. Each time frequency point in Calculate its energy with respect to time With frequency The partial derivatives of are used to obtain the local gradient direction vector at that point. Because the chirp characteristic of LoRa signals exhibits a linear frequency variation over time, the effective signal region on the time-frequency graph will display a constant gradient direction, the slope of which is... Strictly determined by the spreading factor and bandwidth, i.e. The system statistically analyzes the gradient direction distribution of the effective signal region in all samples and constructs a probability density function with the slope as the independent variable. The function exhibits a peak at the slope value corresponding to the standard LoRa configuration, while tending to be uniformly distributed in non-signal regions.

[0033] Based on the above statistical patterns, the system constructs a linearity statistical kernel. The core is initialized to a size of The weight matrix has a central element with a positive weight, and the weights of surrounding elements are assigned based on the deviation of the slope of the line connecting the surrounding element to the central point from the standard Chirp slope. Specifically, if the slope of the line connecting a surrounding element to the central point is close to the standard Chirp slope, the weight of the surrounding element is assigned a positive value. If the peak slope is positive, a positive weight is assigned; conversely, if the deviation exceeds a preset threshold such as 15%, a negative weight is assigned. The initialization matrix is ​​then fine-tuned using an adversarial training mechanism: the generated Chirp signal samples are used as positive examples, and Gaussian white noise samples as negative examples. These are input into a discriminator network containing the convolutional kernel. By minimizing the loss function of the discrimination error, the matrix weights are updated using the backpropagation algorithm until the network's response strength to positive examples is higher than that to negative examples, for example, the signal-to-noise ratio gain exceeds 10dB. The finally converged weight matrix is ​​then established as the linearity statistical kernel of this invention, used to perform physical feature-based gating operations in generative adversarial networks. The construction adopts a discretized slope probability density function. Direct numerical mapping, definition Convolution kernel center element coordinates Weight Initially The remaining eight neighboring elements within the core satisfy Calculate the slope of the geometric line connecting the relative centers. Query the pre-built LoRa signal spreading factor slope statistical distribution table to obtain the corresponding slope value. Standard noiseless signal set probability density And assign the value to the corresponding position weight. For statistical probability Below Weights corresponding to nonlinear characteristic directions Setting the weights to zero creates a sparse weight matrix; during training, only non-zero elements are allowed in the matrix. Within the floating range, updates maintain physical topology constraints; the generator performs slope attention-gated computation while ensuring the Hadamard product tensor dimension alignment, with the input first-order difference tensor dimension... Perform edge padding, filling both ends of the time and frequency axes with unit-width zero-value vectors, expanding the tensor dimension to... Apply the above Linearity statistical kernel convolution, with a stride of 1 and a constant number of channels, reduces the output dimension to [value missing]. A linearity confidence plot is generated, and a spatial coordinate mapping is established between the confidence plot and the original high-dimensional feature tensor. A gating operator is used to suppress nonlinear noise components corresponding to the time step and frequency index.

[0034] Example 5: Before actually deploying the LoRa signal recognition and analysis method of the present invention in a monitoring area with a new environment or unknown electromagnetic background, in order to ensure the recognition stability and parameter analysis accuracy of the system under diverse operating conditions, a standardized on-site deployment pre-calibration procedure needs to be performed. In a silent state without UAV signal transmission, the background noise baseline calibration program of the system is started, and the software radio peripheral is controlled to perform long-period spectrum scanning in the entire monitoring frequency band. The sampling time is set to more than 10 minutes to capture the steady-state background noise and transient pulse interference characteristics in the environment. The system calculates the average noise power spectral density at each frequency point through statistical analysis and uses it to construct a dynamic noise benchmark model. This model is loaded into the front-end preprocessing module as a reference baseline for the adaptive denoising algorithm, ensuring that the influence of environmental noise fluctuations on the signal-to-noise ratio estimation can be compensated in real time during actual operation, so that the subsequent neural network input features are maintained within a relatively stable signal-to-noise ratio range.

[0035] To address potential co-channel interference or multipath effects within a specific area, a field fine-tuning procedure for signal feature templates is implemented. A reference signal source transmits a standard LoRa test signal with known parameters at a typical location within the monitoring area. The receiving system collects these test signals and compares them with preset standard linear frequency modulation texture features to calculate the distortion degree of the time-frequency texture caused by the channel impulse response under the current environment. Based on this distortion parameter, the system uses an online transfer learning algorithm to make minor updates to the input layer weights of a lightweight convolutional neural network to adapt to the channel characteristics unique to the current environment. Simultaneously, the frequency offset and time delay jitter range of the reference signal during transmission are recorded, and the tolerance threshold in the parameter inversion module is updated to ensure that the extraction of bandwidth and symbol period can still converge to the correct values ​​in scenarios with severe multipath or Doppler frequency shift. Through this pre-calibration process, the system completes the feature adaptation to the deployment environment.

[0036] Example 6: To address the uncertainties in parameter calibration and ambiguities in algorithm implementation that the LoRa signal recognition and parsing system involved in this invention may face during actual deployment, this example provides a standardized pre-deployment calibration and model building procedure. It sets the initialization parameters for the linearity statistical kernel, executes an offline optimization parameter-finding procedure, and uses a high-precision arbitrary waveform generator to construct a standard signal sample library containing full-coverage spreading factors SF7-SF12 and full-bandwidth configurations of 125kHz-500kHz. Gaussian white noise with different signal-to-noise ratio levels (-20dB to 0dB) and a step size of 2dB is superimposed. For each signal configuration combination in the sample library, the gradient descent algorithm is used to traverse the parameter space of the linearity statistical kernel's weight matrix. The objective function is to maximize the reconstructed signal-to-noise ratio of the adversarial network output. The optimal weight matrix benchmark value under this configuration is determined. Through this process, a linearity statistical kernel benchmark parameter lookup table indexed by spreading factor, bandwidth, and estimated signal-to-noise ratio is established and stored in the system's non-volatile memory.

[0037] To address environmental adaptability issues in actual operation, an adaptive calibration procedure is implemented for on-site deployment. After power-on, the system automatically enters environmental awareness mode, controlling the RF front-end to collect no less than 100 sets of environmental noise floor data during periods without target signals. These data are used to calculate the spectral flatness and impulse interference distribution characteristics of the current electromagnetic environment, and accordingly dynamically adjust the decision threshold of the discriminator and the noise input distribution parameters of the generator in the generative adversarial network. Simultaneously, the system uses a built-in self-test signal source to transmit a set of test signals containing a specific preamble sequence. The system monitors the characteristic distortion of the signal after passing through the RF front-end and digital processing link through a closed-loop receiving link, calculates and applies a correction coefficient matrix to pre-distort the input feature tensor to eliminate the impact of hardware channel differences on feature extraction accuracy.

[0038] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying and parsing communication signals for LoRa unmanned aerial vehicles, characterized in that, Includes the following steps: Step S101: Obtain discrete radio signal data within the monitoring area, perform short-time Fourier transform on the discrete radio signal data to generate a time-frequency two-dimensional spectrum, and use a convolutional neural network to identify texture features in the two-dimensional spectrum and extract the noisy IQ time series that is determined to be a LoRa signal. Step S102: The truncated noisy IQ time series is mapped to a high-dimensional feature tensor through a one-dimensional convolutional layer, and the high-dimensional feature tensor is input into the generator in the slope statistical attention generative adversarial network. In step S103, the generator performs a gated computation based on statistical attention on the high-dimensional feature tensor according to the physical statistical laws of the linear frequency modulated signal: the gradient of the high-dimensional feature tensor in the time dimension is calculated to generate a first-order difference tensor; a convolution operation is performed on the first-order difference tensor using a preset linearity statistical kernel to generate a linearity confidence map representing the consistency of the frequency change rate. The linearity statistical kernel is a weight matrix constructed based on the slope distribution statistical features of the noiseless LoRa signal samples; then, the linearity confidence map is mapped to a binary attention mask using an activation function; finally, the Hadamard product of the high-dimensional feature tensor and the attention mask is calculated, and a gated feature tensor with nonlinear noise features filtered out is output. Step S104: Input the gated feature tensor into the bidirectional long short-term memory network to reconstruct the temporal phase dependency of the signal, and map the features back to the dimensional numerical space through the transposed convolutional layer to output the denoised IQ time series. Step S105: Perform fast Fourier transform and autocorrelation analysis on the denoised IQ time series to extract bandwidth and symbol period parameters, and calculate the spreading factor based on the product relationship between bandwidth and symbol period.

2. The method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The process of constructing the linearity statistical kernel involved in step S103 includes: constructing a large data training set of noiseless LoRa signals containing multiple spreading factors and bandwidth configurations; extracting the local gradient features of each signal sample in the time and frequency domains in the large data training set, and calculating the statistical distribution probability density of the local gradient features; optimizing the weight matrix using an adversarial training mechanism so that the weight matrix produces a high response value to the gradient features that conform to the statistical distribution probability density, and establishing the converged weight matrix as the linearity statistical kernel.

3. The method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The specific operations for calculating the short-time Fourier transform spectrum in step S101 include: dividing the discrete data of the radio signal into multiple overlapping time windows; performing a fast Fourier transform on each time window and calculating the square of the amplitude spectrum to obtain the power spectral density; and stitching the power spectral densities of multiple time windows on the time axis to generate a two-dimensional grayscale image reflecting the time-frequency energy distribution of the signal.

4. The method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The specific operations of identifying and extracting noisy IQ time series using convolutional neural networks in step S101 include: inputting a two-dimensional grayscale image into a lightweight convolutional neural network, extracting slope texture features in the image using multi-layer convolutional kernels; outputting the probability value of LoRa signal in the current time window through a fully connected layer; when the probability value exceeds a preset decision threshold, backtracking and extracting the corresponding original noisy IQ time series according to the index position of the time window in the original data stream.

5. The method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The specific operations in step S104 for reconstructing the temporal phase dependency using a bidirectional long short-term memory network include: inputting the gated feature tensor into the forward and backward layers of the bidirectional long short-term memory network step by step; using the forward layer to capture the evolution trend of the signal phase and the backward layer to capture the correlation dependency between signal symbols; concatenating the hidden state vectors of the forward and backward layers, and mapping the concatenated vector back to the 1D time domain space through a transposed convolutional layer; restoring the signal amplitude using a denormalization operation to generate a denoised IQ time series.

6. The method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The specific operations for extracting bandwidth and symbol period parameters in step S105 include: performing a fast Fourier transform on the denoised IQ time series to obtain the spectrum, searching for the maximum and second maximum energy points in the spectrum; determining the upper and lower boundary frequencies of the main lobe of the spectrum based on the energy attenuation threshold, and calculating the difference between the upper and lower boundary frequencies as the bandwidth parameter; performing a sliding window autocorrelation operation on the denoised IQ time series, detecting the position of the first peak in the autocorrelation curve, and establishing the time delay corresponding to the position as the symbol period parameter.

7. The method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The specific operations for calculating the spreading factor in step S105 include: multiplying the bandwidth parameter by the symbol period parameter to obtain the number of chips in each symbol period; calculating the base-2 logarithm of the number of chips and rounding the result to obtain the spreading factor.

8. The method for identifying and parsing communication signals for LoRa drones according to claim 7, characterized in that, The method also includes a data-level parameter closed-loop verification step: mapping the calculated center frequency, bandwidth, and spreading factor parameters to register configuration instructions defined by the LoRa communication protocol; and reading the status flag data returned by the LoRa communication module after executing the channel activity detection mode. If the status flag indicates a successful preamble match, then the parameter is established as the final parsing result; If the status flag data indicates a failure to match, the spreading factor is fine-tuned by a preset step size, and a new register configuration instruction is generated to repeat the verification process.

9. The method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The training process of the slope statistical attention generative adversarial network uses a composite loss function for parameter updates, which is defined as follows: ,in, The denoised signal output by the generator. For true noise-free signals, This indicates the Fourier transform operation. To counteract the loss, This represents the mean square error term. Represents the spectrum consistency loss term. and To balance the weights, a hyperparameter is used; the energy distribution of the generated signal in the frequency domain is constrained to be consistent with that of the real signal by a spectral consistency loss term.

10. A method for identifying and parsing communication signals for LoRa drones according to claim 1, characterized in that, The method is deployed as an embedded operator in the edge computing unit of the UAV countermeasure system. The edge computing unit generates and adjusts the modulation parameters of the interference signal based on the spreading factor, and performs data blocking operation on the communication link of the target UAV. The discrete data of radio signals in the monitoring area comes from the real-time acquisition stream of the software radio peripheral connected to the edge computing unit.

Images