A method for dynamically optimizing a detection threshold of a radio frequency signal

Abstract

The application discloses a kind of detection threshold dynamic optimization methods of radio frequency signal, and relates to a kind of detection threshold dynamic optimization methods of radio frequency signal, and its technical scheme points are: obtaining the radio frequency signal received by radio frequency receiving link and real-time temperature of synchronous collection;Radio frequency signal is carried out time domain check and image domain identification and fusion, determine effective pulse signal, pulse confidence and interference intensity;Based on effective pulse signal, signal and noise are separated and extracted, and instantaneous signal-to-noise ratio is generated;Based on pulse confidence, instantaneous signal-to-noise ratio sequence is repaired with abnormal value, and real-time signal-to-noise ratio is determined in combination with dynamic double window and attention timing prediction model;Based on real-time signal-to-noise ratio, interference intensity, real-time temperature and cumulative working length, the target detection threshold is generated by layered neural network optimized by particle swarm algorithm.This application realizes the adaptive optimization of detection threshold under complex working condition.

Description

Technical Field

[0001] This invention relates to the field of radio frequency signal processing, and more specifically, to a method for dynamically optimizing the detection threshold of radio frequency signals. Background Technology

[0002] Automatic Dependent Surveillance-Broadcast (ADS-B) and civil aviation Mode S transponders rely on the 1030MHz interrogation and 1090MHz response frequency bands to broadcast aircraft position, altitude, and identity information. They are core wireless sensing devices for air traffic control surveillance and UAV collision avoidance systems.

[0003] In radio frequency (RF) receiving systems, the setting of the detection threshold directly affects the ability to acquire weak signals and suppress noise interference. Existing solutions typically use a fixed threshold or simply adjust it based on the received signal power, failing to fully consider multiple coupling factors such as real-time changes in signal-to-noise ratio (SNR), multipath interference, temperature drift, and equipment aging. When the SNR decreases due to distance attenuation or electromagnetic interference in the receiving link, a fixed threshold can easily lead to missed detections of valid pulses; while in high SNR or strong clutter scenarios, it may introduce false responses. Furthermore, traditional pulse recognition algorithms lack quantitative assessment of signal confidence in complex environments, SNR estimation relies on a fixed window, and outlier handling strategies are simplistic, making it difficult to handle both abrupt and gradual changes, resulting in delayed or oscillating threshold adjustments. In addition, existing systems lack closed-loop feedback mechanisms between processing modules, failing to autonomously correct parameters based on actual execution status, leading to continuous performance degradation over long-term operation.

[0004] Therefore, there is an urgent need for a radio frequency signal processing method that can adaptively optimize the detection threshold based on real-time signal-to-noise ratio and environmental conditions, and has closed-loop correction capability. Summary of the Invention

[0005] To address the problems of insufficient adaptive detection threshold, poor pulse recognition robustness, and lack of closed-loop optimization mechanism in existing radio frequency receiving systems under complex operating conditions, this invention provides a dynamic optimization method for the detection threshold of radio frequency signals, which achieves high-precision pulse detection, stable signal-to-noise ratio estimation, and adaptive optimal threshold control in all scenarios.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0007] This invention provides a method for dynamically optimizing the detection threshold of radio frequency signals, the method comprising:

[0008] Acquire the radio frequency signal received by the radio frequency receiving link and the real-time temperature collected synchronously;

[0009] The radio frequency signal is subjected to time-domain verification and image-domain recognition, and the time-domain verification results and image-domain recognition results are fused to determine the effective pulse signal, the pulse confidence of the effective pulse signal, and the interference intensity.

[0010] Based on the effective pulse signal, the effective signal and noise are separated and extracted to obtain the effective signal power and noise power, and the instantaneous signal-to-noise ratio is generated by combining the effective signal power and noise power.

[0011] Based on the pulse confidence, outlier repair is performed on the instantaneous signal-to-noise ratio sequence composed of the instantaneous signal-to-noise ratio, and the real-time signal-to-noise ratio is determined by combining the pre-built dynamic dual-window and attention-based temporal prediction model.

[0012] Based on the real-time signal-to-noise ratio, interference intensity, real-time temperature, and cumulative working time, a target detection threshold is generated using a hierarchical neural network optimized by the particle swarm optimization algorithm.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] This invention integrates temporal verification and image domain recognition to quantify pulse confidence and interference intensity, providing quality priors and interference dimension inputs for subsequent signal-to-noise ratio (SNR) processing. Instantaneous SNR is generated by separating and extracting the signal window and noise window based on the effective pulse signal. Differential anomaly repair is performed on the instantaneous SNR based on pulse confidence. Combined with dynamic dual-window adjustment and attention timing prediction, real-time SNR estimation remains smooth and stable under both abrupt and gradual changes. A layered neural network optimized by particle swarm optimization decouples the SNR base mapping from multi-dimensional compensation for interference, temperature, and aging. Network parameters are optimized with the joint minimization of false detection rate, false response rate, and power consumption as the objective, enabling the target detection threshold to adapt to the entire SNR range, wide temperature range, and the changing operating conditions throughout the device's lifecycle. Based on the target detection threshold, detection control is performed on the RF receiving link, and the status is fed back to the front-end processing stage, forming a closed-loop feedback mechanism that enables the system to have autonomous iterative optimization capabilities. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0016] Figure 1 This is a flowchart illustrating a method for dynamically optimizing the detection threshold of radio frequency signals according to an embodiment of the present invention.

[0017] Figure 2 This is a flowchart of pulse recognition and detection provided in an embodiment of the present invention;

[0018] Figure 3This is a flowchart of real-time SNR estimation provided in an embodiment of the present invention;

[0019] Figure 4 This is a flowchart of dynamic optimization of the MTL threshold provided in an embodiment of the present invention;

[0020] Figure 5 This is a simulation diagram of SNR estimation provided in an embodiment of the present invention;

[0021] Figure 6 This is a simulation diagram of SNR estimation provided by existing technology;

[0022] Figure 7 This is the MTL threshold dynamic optimization diagram provided in the embodiments of the present invention. Detailed Implementation

[0023] 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 embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0024] It should be noted that the terms "comprising" or "may include" used in the various embodiments of this application indicate the presence of the claimed function, operation, or element, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in the various embodiments of this application, the terms "comprising," "having," and their cognates are intended only to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or adding one or more combinations of the foregoing.

[0025] It should be understood that terms such as "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0026] Automatic Dependent Surveillance-Broadcast (ADS-B) and civil aviation Mode S transponders rely on the 1030MHz interrogation and 1090MHz response frequency bands to broadcast aircraft position, altitude, and identity information. They are core wireless sensing devices for air traffic control surveillance and UAV collision avoidance systems. The pulse recognition sensitivity, minimum detectable threshold (MTL) adaptive adjustment capability, and signal-to-noise ratio (SNR) estimation accuracy of the receiving link directly determine the long-range target acquisition capability, decoding reliability in low SNR environments, and the long-term operational stability of the entire system. Currently, civil aviation radio frequency receiving systems generally adopt a discrete signal processing architecture, with pulse detection, SNR estimation, detection threshold adjustment, and radio frequency gain compensation modules operating independently. This results in multiple inherent defects under real airborne conditions such as long-range weak signals, multipath reflections, alternating high and low temperatures, and long-term equipment aging. Existing technical solutions struggle to achieve balanced performance across all scenarios.

[0027] In terms of MTL (Mean Transmission Level) detection threshold control, traditional equipment generally uses factory-fixed thresholds to complete pulse decisions. When aircraft fly long distances and transmission losses increase, leading to a significant drop in received SNR (Signal Response Rate), the fixed threshold filters out effective weak pulses, resulting in missed target detection. In low-altitude multi-building and mountainous multipath scenarios, background noise increases, and the fixed threshold may misinterpret clutter as interrogation signals, causing a surge in false responses and significantly increasing the pressure on air traffic control backend data processing. Existing dynamic threshold schemes rely solely on coarsely adjusting the threshold based on received signal power, without quantifying real-time SNR as the adjustment basis. When SNR changes rapidly, threshold adjustment exhibits a significant lag, and detection decision oscillations are prone to occur at the critical point of signal-to-noise ratio, resulting in poor system robustness. Meanwhile, the existing threshold model does not introduce an RF link aging compensation mechanism. Long-term vibration and temperature changes of airborne equipment will cause the gain of the low-noise amplifier to decrease, the connector insertion loss to increase, and the frequency drift of the filter to drift. After accumulating thousands of hours of operation, the system sensitivity will continue to decline and the false detection rate will increase significantly. Moreover, the traditional optimization logic only pursues the maximization of detection accuracy without establishing a power consumption constraint target. Simply increasing the receiver gain to compensate for signal loss will lead to excessive power consumption of the whole machine, which is difficult to adapt to the lightweight and low power supply requirements of airborne equipment.

[0028] At the pulse recognition algorithm level, mainstream detection methods rely on single time-domain cross-correlation or fixed CFAR thresholds to determine pulses, lacking multipath signal compensation logic. Multiple reflection paths in urban airspace cause pulse waveform distortion and peak splitting, resulting in a pulse recognition accuracy of less than 65% in low signal-to-noise ratio scenarios and persistently high message decoding failure rates. Airborne equipment operates in temperatures ranging from -55℃ to 85℃, and the gain and noise floor of RF devices drift significantly with temperature. Traditional Bayesian amplitude verification uses fixed prior parameters, rendering the pulse amplitude determination standard ineffective under extreme high and low temperatures, doubling the probability of false detections and missed detections. Existing solutions only extract pulse time-domain amplitude and edge features for recognition, without introducing image texture dimensions for dual verification. Electromagnetic interference and adjacent channel clutter easily generate false pulses. Furthermore, complex convolutional networks have large parameter counts and high computational latency, making parallel real-time deployment on low-cost, small FPGAs impossible, resulting in significant engineering implementation challenges.

[0029] In the real-time SNR estimation stage, traditional algorithms use fixed-length sliding windows of varying lengths, which cannot match the dynamic characteristics of SNR changes: fixed long windows result in large response delays in scenarios with sudden SNR fluctuations, while fixed short windows cause significant jitter in scenarios with slow attenuation over long distances. Abnormal noise samples are simply removed using the ±3σ criterion without differentiated interpolation based on pulse recognition confidence. Distorted SNR data from low-confidence pulses continuously contaminates the threshold calculation results. Conventional mean filtering fails to uncover the temporal variation patterns of SNR and lacks an attention-based temporal prediction mechanism. In scenarios with rapid SNR jumps caused by high-speed maneuvering aircraft, estimation errors can exceed 0.8dB, directly impacting the MTL threshold and degrading detection performance.

[0030] From a system architecture perspective, current technical solutions all adopt a unidirectional data flow design. Pulse recognition, SNR estimation, threshold execution, and RF gain compensation lack closed-loop feedback links. Threshold execution errors, RF gain adjustment effects, and long-term temperature drift data cannot be used to correct the front-end signal processing algorithm in reverse, and the algorithm lacks the ability to independently iterate and optimize.

[0031] In existing technologies, MTL thresholds are mostly set with fixed values ​​or simply adjusted based on signal power, which has the following drawbacks:

[0032] The existing technical solution is that when the receiving link is affected by distance attenuation and electromagnetic interference, which reduces the SNR, the fixed threshold is prone to causing weak signals to be missed; when the SNR increases, the fixed threshold will introduce noise interference and increase the false response rate.

[0033] Existing technical solutions adjust the threshold only by signal power fluctuations, without combining SNR to quantify signal quality, resulting in lag in threshold adjustment and significant jitter, especially at the SNR critical point where decision oscillations are prone to occur.

[0034] Traditional threshold adjustment algorithms often use a single window to estimate signal characteristics, which cannot take into account both sudden SNR changes (such as sudden interference) and slow fluctuations (such as gradual distance changes), making it difficult to meet the working requirements of full temperature range and wide dynamic range.

[0035] Therefore, there is an urgent need for a dynamic optimization scheme for MTL threshold based on real-time SNR feedback, with accurate and stable algorithms, and adaptability to complex scenarios. To address this, this invention provides a dynamic optimization method for the detection threshold of radio frequency signals. By designing a pulse recognition optimization module, a dual-window-Attention-LSTM fusion estimation algorithm, a PSO-hierarchical neural network adaptive MTL mapping algorithm, and a closed-loop feedback signal processing flow, it simultaneously realizes pulse signal recognition, real-time SNR estimation, dynamic optimization of MTL threshold, link aging compensation, and power consumption balancing. This improves the detection accuracy of weak signals and the reliability of pulse recognition, reduces the missed detection rate, false detection rate, and pulse recognition error rate, and is adaptable to complex radio frequency communication scenarios such as air traffic control transponders, radar receivers, and satellite radio frequency receivers.

[0036] Figure 1 This is a flowchart illustrating a method for dynamically optimizing the detection threshold of radio frequency signals according to an embodiment of the present invention. This method is applicable to radio frequency systems such as air traffic control transponders, radar receivers, and satellite radio frequency receivers that need to maintain stable detection performance in complex electromagnetic environments. The method includes the following steps:

[0037] S101, acquire the radio frequency signal received by the radio frequency receiving link and the real-time temperature collected synchronously.

[0038] In this embodiment, an antenna and RF front-end circuit receive RF signals in the 1030MHz band in the airspace. The 1030MHz band signals acquired in the airspace are filtered and denoised to remove out-of-band interference while retaining the characteristics of the interrogation pulse signal. Simultaneously, a temperature sensor integrated into the RF receiving link synchronously acquires real-time temperature data of the current environment. This synchronous acquisition of RF signals and real-time temperature ensures that subsequent processing stages can obtain complete input information reflecting the current channel state and equipment operating conditions. It should be understood that the method of acquiring RF signals is not limited to a specific antenna type or RF front-end topology. For example, a superheterodyne receiver architecture, a zero-IF receiver architecture, or a direct sampling receiver architecture can be used, as long as a pre-processed digital RF signal can be output. Real-time temperature acquisition can also be achieved using different types of temperature sensors, such as thermistors, digital temperature sensors, or thermocouples.

[0039] S102 performs time-domain verification and image-domain recognition on the radio frequency signal, and fuses the time-domain verification results and image-domain recognition results to determine the effective pulse signal, the pulse confidence of the effective pulse signal, and the interference intensity.

[0040] In this embodiment, Figure 2This is a flowchart of pulse recognition and detection provided in an embodiment of the present invention; as shown below. Figure 2 As shown, radio frequency (RF) signals are identified from two parallel dimensions. The time-domain verification dimension extracts amplitude matching features and statistical significance of the RF signal on the time axis sequentially through normalized autocorrelation detection, adaptive CFAR, temperature-adaptive Bayesian verification, and pulse slope verification, outputting the time-domain verification result. The image-domain recognition dimension maps the one-dimensional time-domain waveform to a two-dimensional pixel matrix using a CNN image recognition model, and extracts the spatial texture features of the signal on the time-amplitude plane using a convolutional neural network, outputting the image-domain recognition result. The two results are weighted and fused to determine not only whether the current signal is a valid pulse, but also to quantify the pulse confidence and interference intensity. The pulse confidence reflects the reliability of the current recognition result, providing a priori basis for subsequent differential processing of signal-to-noise ratio data; the interference intensity provides the interference dimension input for subsequent threshold optimization by quantizing clutter power.

[0041] This embodiment further performs time-domain verification and image-domain recognition on the radio frequency signal, and fuses the time-domain verification results and image-domain recognition results to determine the effective pulse signal, the pulse confidence of the effective pulse signal, and the interference intensity. The specific implementation method is described in detail below:

[0042] The radio frequency signal is sequentially subjected to normalized autocorrelation pulse detection, adaptive CFAR noise floor suppression, temperature-adaptive Bayesian amplitude verification, and pulse slope verification to determine the time-domain verification confidence probability. Specifically, the time-domain verification constitutes a multi-level cascaded signal screening link, with each level verifying candidate pulses from different dimensions, progressively improving the reliability of the identification results.

[0043] The autocorrelation calculation logic is optimized by introducing a noise variance suppression factor and a multipath delay compensation term. Normalized autocorrelation impulse detection is then performed based on the optimized logic. The normalized autocorrelation detection calculates the autocorrelation value using the following formula:

[0044] In this context, x(n) is the pulse signal sequence of the real-time acquired radio frequency signal, k is the delay step size, N is the signal sequence length, μx is the mean of x(n), σn² is the background noise variance (estimated in real-time by adaptive CFAR), ∂ is the noise variance suppression factor, used to enhance the sharpness of the autocorrelation peak in low signal-to-noise ratio scenarios, λ is the multipath compensation coefficient, and τ is the multipath delay. The multipath delay compensation term is determined based on the multipath delay calculated from the additional multipath propagation distance of the radio frequency signal, specifically calculated using the formula τ=2d / c, where d is the additional multipath propagation distance, c is the speed of light, and τ is estimated in real-time through the pulse phase difference within the signal window. When the multipath interference intensity is greater than or equal to a preset interference threshold, the weight of the multipath delay compensation term is increased. For example, when the multipath interference intensity is greater than or equal to 3dB, the weight of λ is automatically increased from 0.3 to 0.7. When the normalized autocorrelation value RINCC(k) is greater than or equal to the adaptive CFAR threshold, it is determined to be a potential pulse signal. It should be understood that the specific value of the noise variance suppression factor is not limited to the above form. It can also adopt an adaptive adjustment method based on the noise power spectral density or a lookup method based on the signal-to-noise ratio estimate. The weight adjustment strategy of the multipath delay compensation term can also be recalibrated according to the statistical characteristics of the actual multipath environment.

[0045] As one implementation method, adaptive CFAR noise suppression includes: estimating background noise power using a cell-average constant false alarm rate algorithm, and calculating a dynamic detection threshold based on the average signal power of reference cells, a preset false alarm rate, and the number of reference cells. The threshold calculation formula for adaptive CFAR is: ,in, The average signal power of M reference cells is obtained through the formula... Calculate P fa The preset false alarm rate is defined by M, where M is the number of reference units. In this embodiment, P... fa The value is 10⁻ 6 M is set to 32 to adapt to the pulse characteristics of the 1030MHz signal. This dynamic detection threshold can adaptively adjust to follow the fluctuations in background noise. It should be understood that the CFAR algorithm is not limited to the Cell Average Constant False Alarm Rate (CA-CFAR), but can also adopt variants such as Ordered Statistical Constant False Alarm Rate (OS-CFAR) or Minimum Selective Constant False Alarm Rate (SO-CFAR).

[0046] As one implementation method, temperature-adaptive Bayesian amplitude verification includes: dynamically adjusting the prior parameters of the Bayesian verification based on real-time temperature. The core formula of temperature-adaptive Bayesian verification includes:

[0047] Temperature-adaptive Bayesian verification formula, i.e. ;

[0048] Temperature-adaptive Bayesian variance formula, i.e. ;

[0049] and the Bayesian amplitude estimation formula Where T is the real-time temperature (°C). =10dBm =0.8dBm² is the prior parameter at room temperature (25℃). =0.005dBm / ℃ =0.01dBm² / ℃ is the temperature correction factor, which is determined based on the temperature difference between the real-time temperature and the ambient temperature reference. To measure the pulse amplitude, =0.5dBm² represents the sampling noise variance, and Ast is the pulse amplitude estimate after temperature adaptive calibration. By adjusting the prior parameters through temperature adaptation, the pulse amplitude calibration error can be reduced from ±0.8dB to ±0.2dB under extreme temperatures. It should be understood that the determination of the temperature correction coefficient is not limited to the above linear model; methods such as polynomial fitting based on experimental calibration or piecewise linear interpolation can also be used.

[0050] Pulse slope verification includes: calculating the rising edge slope and falling edge slope of the pulse. When the rising edge slope is greater than or equal to the rising edge slope threshold and the falling edge slope is greater than or equal to the falling edge slope threshold, it is determined to be a valid potential pulse.

[0051] Calculate the pulse rise slope krise = ΔA / Δt and fall slope kfall = ΔA / Δt. Taking a 1030MHz interrogation pulse as an example, the corresponding rise slope threshold can be set to 20dBm / μs, and the fall slope threshold can be set to 12.5dBm / μs. Only when k rise ≥20dBm / μs and k fall A potential pulse with a value ≥12.5 dBm / μs is considered valid; otherwise, it is considered spurious interference due to slow rising edges caused by electromagnetic clutter and is discarded. It should be understood that the specific value of the slope threshold is not limited to the above example and can be reset according to the pulse specifications of the target signal.

[0052] After the above four-level time-domain verification, the normalized autocorrelation detection, adaptive CFAR, temperature-adaptive Bayesian verification, and impulse slope verification are combined to output the time-domain verification confidence probability P. conf .

[0053] In a further example, the time-domain waveform of the radio frequency signal is mapped to a two-dimensional pixel matrix, and features are extracted from the two-dimensional pixel matrix using a CNN image recognition model to determine the image recognition confidence probability.

[0054] Specifically, using time as the horizontal axis and amplitude as the vertical axis, the time-domain waveform of the radio frequency signal is quantized into grayscale values ​​to generate a two-dimensional pixel matrix. The horizontal axis covers sampling points of a preset signal period, and the vertical axis corresponds to the quantization range of the signal power. For example, for an interrogation signal in the 1030MHz band, the horizontal axis can be set to cover 100 sampling points, corresponding to a preamble pulse of approximately 3.5μs and the start segment of subsequent symbols; the vertical axis linearly maps the signal power to grayscale levels from 0 to 255, generating a 32×32 pixel pulse time-domain grayscale map.

[0055] It should be understood that the dimensions and grayscale methods of the two-dimensional pixel matrix are not limited to the examples above. They can also be adjusted to other sizes (such as 64×64 or 16×16) depending on the signal bandwidth and sampling rate. Grayscale can also employ non-linear mapping to enhance the contrast of weak signal areas.

[0056] The CNN image recognition model comprises an input layer, at least one convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer receives a two-dimensional pixel matrix, which is then processed through at least one convolutional layer to extract local spatial features. The pooling layer performs max pooling to reduce the dimensionality of these local spatial features, yielding image texture features. The fully connected layer performs classification mapping on the image texture features, and the output layer outputs the classification probabilities of valid and false pulses. The classification probability of valid pulses is used as the image recognition confidence probability. As a preferred implementation, this model employs a lightweight architecture: the input layer receives a 32×32 pixel grayscale image; the convolutional layer uses two sets of 3×3 convolutional kernels, with 8 and 16 kernels respectively, employing the ReLU activation function; the pooling layer uses 2×2 max pooling; and the fully connected layer maps the extracted texture features to the two-dimensional output space. The total number of parameters in the entire network is controlled to within 80, the inference latency is no more than 0.2 ns, and it can be deployed in parallel on embedded hardware platforms such as FPGAs. Of course, in other embodiments, the number of convolutional layers, the size of the convolutional kernel, and the type of activation function can also be replaced according to the actual computing power and recognition accuracy requirements. For example, activation functions such as LeakyReLU or ELU can be used, or batch normalization layers can be added to accelerate training convergence.

[0057] After obtaining the time-domain verification confidence probability and the image recognition confidence probability, the fusion weight of the time-domain verification confidence probability and the image recognition confidence probability is dynamically adjusted based on the signal-to-noise ratio status.

[0058] Specifically, when the current signal-to-noise ratio is low, the reliability of the time-domain amplitude information decreases, and the system automatically increases the weight of the image recognition confidence probability to maintain recognition performance by taking advantage of the image texture features being insensitive to amplitude fluctuations. Conversely, when the signal-to-noise ratio is high, the time-domain waveform is clear and complete, and the system increases the weight of the time-domain verification confidence probability to take advantage of the characteristics of low computational load and fast response speed in time-domain processing.

[0059] The pulse confidence level is obtained by weighting and fusing the time-domain verification confidence probability and the image recognition confidence probability based on the adjusted fusion weights, and the effective pulse signal is determined based on the pulse confidence level. The clutter power is then quantized synchronously to determine the interference intensity.

[0060] The fusion process follows the formula below: ;in, Indicates the impulse confidence level. This represents the confidence probability for image recognition. This represents the confidence probability of the time-domain verification. and These are the corresponding fusion weights and satisfy the following conditions: + =1. Under normal signal-to-noise ratio conditions. Set it to 0.4. Set to 0.6; when the real-time signal-to-noise ratio falls below the preset threshold and enters the low signal-to-noise ratio range, Automatically increases to 0.5 or even 0.6. The corresponding decrease. When If the signal is greater than or equal to a preset confidence threshold (e.g., 0.85), it is considered a valid pulse signal; otherwise, it is considered false interference and is discarded.

[0061] While outputting the valid pulse signal sequence, the power of clutter components that are not determined to be valid pulses is statistically quantified, and interference intensity information is output. This interference intensity reflects the overall level of clutter and interference in the current electromagnetic environment.

[0062] S103 separates and extracts the effective signal and noise based on the effective pulse signal to obtain the effective signal power and noise power, and combines the effective signal power and noise power to generate the instantaneous signal-to-noise ratio.

[0063] In this embodiment, Figure 3 This is a flowchart of real-time SNR estimation provided in an embodiment of the present invention; as shown below. Figure 3 As shown, based on the effective pulse signal, a signal window and a noise window are divided in the time domain, and the pure effective signal component and the purified noise component are extracted respectively. The time-domain mean square value of the signal within each window is calculated and converted into a power quantization value. The difference between the effective signal power and the noise power is used as the instantaneous signal-to-noise ratio at the current pulse moment. This instantaneous signal-to-noise ratio directly reflects the real-time quality of the current pulse signal.

[0064] This embodiment further details the separation and extraction of effective signal and noise based on the effective pulse signal to obtain the effective signal power and noise power, and provides a detailed explanation of the specific implementation method for generating the instantaneous signal-to-noise ratio in combination with the effective signal power and noise power, as follows:

[0065] Centered on the peak position of the pulse output from the normalized autocorrelation pulse detection (INCC), an interval corresponding to the main lobe width of the pulse is selected as the signal window to extract effective signal components. Specifically, the effective signal components are extracted using the peak position n of the pulse output from the INCC detection. peak Centered on the signal, the interval corresponding to the width of the main lobe of the pulse is selected as the signal window, with a length of Ns. Taking a 1030MHz pulse signal as an example, Ns is taken as 8 to 12 signal cycles, approximately 7.76ns to 11.64ns. Within this interval, the signal consists of pure effective pulse components, denoted as x. s (n), where, It should be understood that the selection of the signal window is not limited to a symmetrical window centered on the pulse peak. An asymmetrical window can also be used based on the asymmetrical characteristics of the pulse waveform, or the offset can be selected based on the starting point of the pulse rising edge.

[0066] The interval outside the signal window is selected as the noise window. Abnormal spikes in the noise signal within the noise window are removed using preset criteria to obtain the purified noise sequence.

[0067] Specifically, the noise window is selected from the interval outside the signal window, with a length of Nn, which is generally not less than 32 signal cycles. For the noise signal x within the noise window... n (n) Abnormal spikes are removed using the ±3σ criterion to avoid the interference of pulse sidelobes in the noise estimation, resulting in the purified noise sequence x. n-clean (n). It should be understood that the criteria for removing abnormal peaks are not limited to the ±3σ criterion, but may also include robust estimation criteria based on median absolute deviation (MAD) or truncation criteria based on percentiles.

[0068] The time-domain mean square value of the effective signal component within the signal window and the time-domain mean square value of the purified noise sequence within the noise window are calculated respectively. The time-domain mean square values ​​and the time-domain mean square values ​​are then converted to obtain the effective signal power and the noise power.

[0069] Specifically, the mean square power method is used to calculate the quantization value. The calculation process for the effective signal power is as follows: first, calculate the time-domain mean square value of the sequence within the signal window. Then convert to dBm units , where P ref For reference power, 1mW is typically used. The noise power is calculated as follows: calculate the time-domain mean square value of the purification sequence within the noise window. Then convert to dBm units It should be understood that power calculation methods are not limited to the time-domain mean square power method; frequency-domain power spectrum integration or power estimation methods based on autocorrelation functions can also be used.

[0070] The instantaneous signal-to-noise ratio is obtained by calculating the difference between the effective signal power and the noise power.

[0071] Specifically, through the formula SNR t =SN calculates the instantaneous signal-to-noise ratio (SNR) at the current pulse moment. This value directly reflects the real-time quality of the current pulse signal. If SNR... t A negative value indicates that the signal is overwhelmed by noise, but due to the CFAR threshold filtering in the pre-pulse recognition stage, such invalid signals have been eliminated, resulting in a final output SNR. t All are valid quantified values.

[0072] S104, based on the pulse confidence, outlier repair is performed on the instantaneous signal-to-noise ratio sequence composed of the instantaneous signal-to-noise ratio, and the real-time signal-to-noise ratio is determined by combining the pre-built dynamic dual-window and attention temporal prediction model.

[0073] In this embodiment, the combination of a confidence-driven repair mechanism and dynamic window fusion estimation ensures that the signal-to-noise ratio data can truly reflect the channel quality.

[0074] This embodiment further details the specific implementation of outlier repair for instantaneous signal-to-noise ratio sequences based on impulse confidence, and the determination of real-time signal-to-noise ratio using a pre-built dynamic dual-window and attention-based temporal prediction model, as follows:

[0075] Based on the confidence interval of the impulse confidence level, a differentiated repair strategy is implemented for abnormal signal-to-noise ratio (SNR) data in the instantaneous SNR sequence. The confidence interval includes high-confidence, medium-confidence, and low-confidence intervals.

[0076] The core of the repair strategy provided in this embodiment lies in using the pulse confidence of the output of the pre-pulse identification as a priori measure of data quality, refining the single anomaly handling logic into a hierarchical processing mechanism, thereby achieving a balance between preserving effective information and suppressing noise pollution.

[0077] When the pulse confidence level is in the high confidence range, cubic spline interpolation is used to repair abnormal signal-to-noise ratio (SNR) data. For example, when the pulse confidence level is greater than or equal to 0.9, the pulse is determined to be a high-confidence sample. Even if the corresponding instantaneous SNR fluctuates beyond ±3σ, it is highly likely caused by physical factors such as fast channel fading rather than spurious interference. The cubic spline interpolation algorithm can preserve the temporal continuity and local trend characteristics of the signal to the greatest extent while repairing outliers, avoiding data breaks caused by simple removal.

[0078] When the pulse confidence level is within the medium confidence range, a linear prediction algorithm is used to correct abnormal SNR data by combining the SNR trend of a preset number of cycles before and after. For example, when the pulse confidence level is between 0.7 and 0.9, the linear prediction algorithm is used to correct outliers by combining the SNR change trend of five cycles before and after. This method provides reasonable estimates based on historical trends, maintaining the overall consistency of the sequence.

[0079] When the pulse confidence level is in the low confidence range, abnormal signal-to-noise ratio (SNR) data are removed. For example, when the pulse confidence level is less than 0.7, it indicates that the pulse is highly likely to be a false interference or severely distorted signal, and its corresponding SNR data has lost its reference value. It is directly removed from the sequence to prevent contamination of subsequent statistical estimations. It should be understood that the thresholds of 0.9, 0.7, etc., mentioned above are only illustrative examples. In practical applications, they can be adaptively adjusted according to the false alarm rate requirements of the RF system and the characteristics of the channel environment. Alternatively, an adaptive threshold calibration method based on the statistical distribution of historical data can be used.

[0080] By combining a pre-built dynamic dual-window and attention-based temporal prediction model, the real-time signal-to-noise ratio is determined, including:

[0081] After outlier repair, the repaired sequence is smoothed and then predicted.

[0082] Calculate the variance of the signal-to-noise ratio (SNR) data within a short-term time window. This variance, Vars, characterizes the degree of fluctuation in the SNR within the current time period and is a key indicator for determining the channel state type.

[0083] Based on the comparison results between the variance and a first preset threshold and a second preset threshold, the lengths of the short-term and long-term time-series windows are dynamically adjusted. Specifically, when the variance is greater than or equal to the first preset threshold, the short-term time-series window is shortened and the long-term time-series window is extended. For example, when Vars ≥ 1 dB², it is determined that the current situation is a scenario with a sudden change in signal-to-noise ratio (SNR), and the short-term window length is shortened from the initial value of 15 signal periods to 10 signal periods to improve response speed, while the long-term window length is extended from the initial value of 75 signal periods to 90 signal periods to provide a more stable background reference benchmark. When the variance is less than the second preset threshold, the short-term time-series window is extended and the long-term time-series window is shortened. For example, when Vars < 0.3 dB², it is determined that the current situation is a scenario with a slow change in SNR, and the short-term window is extended to 20 signal periods to utilize more samples to suppress random jitter, while the long-term window is shortened to 60 signal periods to reduce computational complexity. When the variance is between the two thresholds, the initial window length remains unchanged.

[0084] Based on the adjusted short-term and long-term time-series windows, the short-term mean signal-to-noise ratio (SNR) and the long-term mean SNR are calculated, respectively. These two means reflect the immediate level and long-term trend of the SNR, respectively.

[0085] The short-term signal-to-noise ratio (SNR) mean, long-term SNR mean, and variance are input into the attention-based time-series prediction model. The model calculates the weights of the short-term SNR mean, long-term SNR mean, and variance through the attention mechanism, and outputs the predicted SNR value for the next time step.

[0086] In this embodiment, the attention-based temporal prediction model employs a lightweight Attention-LSTM network structure with a 3-dimensional input, corresponding to the short-term mean, long-term mean, and short-term variance. The number of hidden layer neurons is set to 16 to accommodate the resource constraints of embedded hardware such as FPGAs. The activation function is tanh. The output dimension is 1, representing the predicted signal-to-noise ratio (SNRlstm) for the next time step. The initial learning rate is set to 0.001, and the training iterations are performed for 500 epochs. The attention weights are calculated using a cosine similarity function, enabling the model to automatically focus on the short-term mean and variance in abrupt change scenarios and on the long-term mean in a gradual change scenario. It should be understood that the attention-based temporal prediction model is not limited to the LSTM structure and can also be replaced by temporal prediction networks such as GRU (Gated Recurrent Unit) or Transformer.

[0087] The real-time signal-to-noise ratio (SNR) is determined by weighting and fusing the short-term mean SNR, long-term mean SNR, and predicted SNR based on adaptive weights. The fusion follows the formula below: ,in, This represents the real-time signal-to-noise ratio of the final output. This represents the short-term signal-to-noise ratio mean. This represents the long-term mean signal-to-noise ratio. This represents the predicted signal-to-noise ratio (SNR) output of the attention-based temporal prediction model. α, β, and γ are adaptive fusion weights for the corresponding SNR mean, satisfying α + β + γ = 1. These weights are dynamically generated by the attention mechanism based on the current channel state. Initially, α is set to 0.3, β to 0.5, and γ to 0.2. During moments of abrupt SNR changes, the weights of α and γ automatically increase, allowing the output to quickly follow the actual changes; during moments of stable SNR, the weight of β dominates, keeping the output smooth and stable.

[0088] S105 generates target detection thresholds based on real-time signal-to-noise ratio, interference intensity, real-time temperature, and cumulative working time through a hierarchical neural network optimized by particle swarm optimization.

[0089] Specifically, Figure 4 This is a flowchart of dynamic optimization of the MTL threshold provided in an embodiment of the present invention, such as... Figure 4As shown, the hierarchical neural network consists of two sub-networks: a main network and a compensation network. Its network parameters are optimized offline to jointly minimize the false response rate, false alarm rate, and power consumption. During runtime, the main network outputs a basic detection threshold based on the real-time signal-to-noise ratio (SNR), while the compensation network calculates environmental compensation based on the real-time SNR, interference intensity, and real-time temperature. The two are then combined with a link aging factor for segmented calibration, ultimately generating a target detection threshold adapted to the current operating conditions.

[0090] This embodiment further details the specific implementation of generating a target detection threshold using a hierarchical neural network optimized by the particle swarm optimization algorithm, based on real-time signal-to-noise ratio, interference intensity, real-time temperature, and cumulative working time, as follows:

[0091] The hierarchical neural network consists of a main network and a compensation network; the network parameters of the main network and the compensation network are determined by optimizing the particle swarm optimization algorithm with the goal of jointly minimizing the false response rate, false response rate and power consumption.

[0092] Specifically, the particle encoding corresponds to the weights and biases of the neural network (a total of 18 parameters), the number of particles is set to 20, the number of iterations is set to 30, and the inertia weight is linearly decayed from 0.9 to 0.4 to ensure convergence accuracy. The objective function comprehensively considers three indicators, with weight allocations of 0.5 for false negative rate, 0.3 for false response rate, and 0.2 for power consumption. It should be understood that the parameter configuration of the particle swarm optimization algorithm is not limited to the above example; an adaptive inertia weight strategy or the introduction of mutation operators can also be used to enhance the global search capability; the optimization algorithm itself can also be replaced by other swarm intelligence optimization methods such as genetic algorithms and differential evolution algorithms.

[0093] The real-time signal-to-noise ratio is input into the main network, and the nonlinear relationship between the signal-to-noise ratio and the basic detection threshold is fitted to output the basic detection threshold.

[0094] Specifically, the main network adopts a three-layer fully connected structure. The input layer is one-dimensional (real-time signal-to-noise ratio), the hidden layer has 8 neurons, and the output layer is mapped to a preset threshold range (e.g., -85dBm to -75dBm) via a sigmoid activation function. The total number of parameters is controlled within 100 to adapt to FPGA deployment requirements. It should be understood that the structure of the main network is not limited to a three-layer fully connected network; deep networks with residual connections or radial basis function networks can also be used.

[0095] Real-time signal-to-noise ratio, interference intensity, and real-time temperature are input into the compensation network to determine the environmental compensation amount. The compensation network independently receives three-dimensional feature input and calculates the environmental compensation amount. The formula for calculating the compensation amount is: Where I is the interference intensity and T is the real-time temperature. For interference compensation coefficient, This represents the temperature compensation coefficient. As a preferred example, wint is set to 0.1 dB / dB, and wT is set to 0.02 dB / ℃.

[0096] The initial detection threshold is determined based on the basic detection threshold and the environmental compensation amount.

[0097] Through the formula MTL base =MTL base1 +MTL comp By superimposing the two values, we obtain an initial detection threshold that reflects both the current channel quality and includes environmental corrections.

[0098] The link aging factor is calculated based on the cumulative operating time, and the signal-to-noise ratio (SNR) range in which the real-time SNR falls is determined. The link aging factor calculation unit 640 calculates the link aging factor η based on the device's cumulative operating time t (in hours), using the following formula: .

[0099] Components in the RF receiver link, such as low-noise amplifiers, filters, and connectors, exhibit irreversible aging phenomena such as gain attenuation, increased insertion loss, and frequency drift as their cumulative operating time increases. After 1000 hours of operation, the η value is 1.1, meaning that the threshold needs to be adjusted accordingly to offset approximately 10% performance degradation. It should be understood that the calculation method for the link aging factor is not limited to the linear model described above; exponential attenuation models or piecewise linear models based on experimental aging data can also be used.

[0100] By using the link aging factor and the segmented calibration coefficient corresponding to the signal-to-noise ratio range, the initial detection threshold is corrected to generate the target detection threshold.

[0101] Specifically, the initial detection threshold is calibrated in segments based on the different ranges of the real-time signal-to-noise ratio and the link aging factor.

[0102] The signal-to-noise ratio (SNR) range includes a high SNR range, a medium SNR range, and a low SNR range. When the real-time SNR is in the high SNR range, the initial detection threshold is adjusted downwards using the first segment calibration coefficient and the link aging factor to generate the target detection threshold. For example, when SNR ≥ 10dB, the target detection threshold MTL is... opt =MTL base -k1·SNR-10·η, where k1 is the first segment calibration coefficient, with an example value of 0.25dB / dB. This range focuses on lowering the threshold as much as possible to improve the dynamic range and enhance the ability to capture weak signals while ensuring no false alarms.

[0103] When the real-time signal-to-noise ratio (SNR) is in the medium SNR range, the initial detection threshold is smoothly corrected using the second segmented calibration coefficient and the link aging factor to generate the target detection threshold. For example, when 6dB ≤ SNR < 10dB, the MTL... opt =MTL base +k2·8-SNR·η, where k2 is the second segment calibration coefficient, with an example value of 0.15dB / dB. This interval focuses on a smooth transition to maintain demodulation stability.

[0104] When the real-time signal-to-noise ratio (SNR) is in the low SNR range, the initial detection threshold is adjusted upwards using the third segment calibration coefficient and the link aging factor to generate the target detection threshold. For example, when SNR < 6dB, MTL opt =MTL base +k3·6-SNR·η, where k3 is the third segment calibration coefficient, with an example value of 0.4dB / dB. This interval focuses on moderately raising the threshold to suppress the noise overwhelm effect. It should be understood that the above segmentation formula and coefficients are only one specific implementation method. In practical applications, calibration correction can also be achieved by continuous functions or polynomial fitting, and the threshold for dividing the signal-to-noise ratio interval can also be recalibrated according to the specific application scenario.

[0105] Based on the method described in the above embodiments, the method provided in this embodiment of the invention further includes: performing detection control on the radio frequency receiving link based on a target detection threshold, and correcting the parameters of time-domain verification, image-domain recognition, outlier repair, and hierarchical neural network based on the state feedback of the detection control.

[0106] Specifically, the detection decision threshold of the radio frequency receiving link is adjusted in real time according to the target detection threshold, and the execution status information is simultaneously fed back to the upstream processing module. The execution status information includes the actual trigger frequency of the threshold, the boundary value trigger status, and the detection accuracy deviation.

[0107] When a deviation between the actual performance and the expected result is detected, the relevant parameters are automatically corrected through a closed-loop feedback path. For example, if the feedback data shows a high false alarm rate, the false alarm rate setting of the adaptive CFAR noise suppression in the time domain verification can be adjusted, or the classification threshold of the CNN image recognition model in the image domain recognition can be adjusted; if the feedback data shows a systematic deviation in the signal-to-noise ratio estimation, the confidence interval division threshold or the initial length configuration of the dynamic dual window in outlier repair can be corrected; if the feedback data shows a continuous deviation between the threshold and the actual optimal value, the network weights and segmented calibration coefficients of the hierarchical neural network can be fine-tuned through PSO parameter optimization.

[0108] Every 100 signal cycles, based on the demodulation effect (including false negative and false positive rates) and pulse recognition accuracy, the network parameters and calibration coefficients are fine-tuned using a particle swarm optimization algorithm. Every 100 hours, the temperature correction coefficient and link aging factor η are updated to adapt to link aging and environmental changes. This closed-loop feedback mechanism enables the system to perform autonomous calibration and continuous optimization, maintaining stable detection performance under long-term operating conditions of device aging and environmental changes without the need for regular manual calibration. It should be understood that the triggering cycle and correction strategy for feedback correction are not limited to the above examples. The triggering conditions can also be dynamically adjusted according to the operating status. For example, parameter correction can be triggered immediately when continuous degradation of performance indicators is detected, or an adaptive triggering mechanism based on sliding window statistics of performance indicators can be adopted.

[0109] For example, when the real-time signal-to-noise ratio is in the high signal-to-noise ratio range, the gain of the low-noise amplifier in the RF receiving link is reduced to suppress noise interference caused by strong near-field signals.

[0110] When the real-time signal-to-noise ratio is in the medium signal-to-noise ratio range, maintain the current gain of the low-noise amplifier to keep the demodulation process stable;

[0111] When the real-time signal-to-noise ratio is in the low signal-to-noise ratio range, the low-noise amplifier is triggered to perform gain compensation operation to help improve the link's reception performance for weak signals.

[0112] The actual gain execution state of the low-noise amplifier is acquired synchronously and fed back to the hierarchical neural network to fine-tune the generation parameters of the target detection threshold.

[0113] This embodiment uses an airborne transponder in a civil aviation secondary surveillance radar system as a specific application scenario to verify the performance of the dynamic optimization method for radio frequency signal detection threshold provided in this application. The test object is an interrogation signal in the 1030MHz band. The receiving link includes a cascaded low-noise amplifier (noise figure ≤ 1.2dB, gain range 22dB to 27dB) and an FPGA demodulation module (minimum demodulation SNR ≥ 6dB). The target detection threshold optimization range is set to -85dBm to -75dBm. The test environment covers typical operating conditions such as high altitude and long distance (100km to 370km), low signal-to-noise ratio (0dB to 20dB), and extreme temperature (-55℃ to 70℃).

[0114] The system parameters are configured as follows: In the pulse recognition stage, the number of reference units M for normalized autocorrelation detection is set to 32, and the preset false alarm rate P... fa Set to 10⁻ 6 The allowable amplitude error ΔA is set to 0.5dB, and the prior parameter is set at room temperature. =10dBm =0.8dBm², sampling noise variance =0.5dBm²; The input image size of the CNN image recognition model is 32×32 pixels, the convolution kernel configuration is two sets of 3×3 convolution kernels (8 and 16 respectively), 2×2 max pooling, the total number of network parameters is 78, and the inference latency is 0.18ns. In the signal-to-noise ratio estimation stage, the signal window length Ns is set to 10 signal cycles (approximately 9.7ns), and the noise window length Nn is set to 32 signal cycles (approximately 31.04ns); the initial parameters of the dual windows are set as short window N1 = 15 signal cycles, long window N2 = 75 signal cycles, and the sliding step size is 1 signal cycle; The Attention-LSTM network adopts a 3-layer structure with an input dimension of 3, 16 hidden layer neurons, tanh activation function, learning rate of 0.001, and 500 training iterations. During the threshold optimization phase, the segmented calibration coefficients were set to k1=0.25dB / dB, k2=0.15dB / dB, and k3=0.4dB / dB, respectively; the environmental compensation coefficient was set to w. int =0.1dB / dB, w T =0.02dB / ℃; the link aging factor calculation formula is η=1+0.0001·t (t is in hours); the number of particles in the PSO optimization algorithm is set to 20, the number of iterations is set to 30, and the objective function weight allocation is 0.5 for false detection rate, 0.3 for false response rate, and 0.2 for power consumption.

[0115] like Figure 5 and Figure 6 As shown, regarding pulse recognition performance, under conditions of strong multipath interference and low signal-to-noise ratio, the pulse recognition accuracy of the traditional single time-domain cross-correlation detection method is 62.3%, while the pulse recognition accuracy is improved to 96.8% after adopting the dual-dimensional fusion method of time-domain verification and image-domain recognition provided in this application. In extreme temperature scenarios (-55℃ and 70℃), thanks to the temperature-adaptive Bayesian verification mechanism, the pulse recognition error rate is reduced to 0.8%, which is 81% lower than the traditional fixed prior parameter scheme.

[0116] Regarding signal-to-noise ratio (SNR) estimation performance, in scenarios with sudden SNR fluctuations, the traditional fixed-window scheme exhibits an estimation response delay exceeding 5 signal cycles and significant overshoot. In contrast, the dual-window dynamic adjustment and Attention-LSTM fusion estimation algorithm employed in this application has a response delay of no more than 1 signal cycle, no overshoot, and fluctuation amplitude controlled within 0.08 dB. In a low SNR steady-state scenario (SNRt = 5 dB), the SNR estimate is 4.88 dB, with an absolute estimation error ≤ 0.12 dB. In full-temperature range testing (-55℃ to 70℃), the SNR estimation stability remains within ±0.12 dB, representing an 88% improvement in accuracy compared to the traditional fixed-window scheme.

[0117] like Figure 7As shown, in terms of detection threshold optimization performance, the detection threshold curve generated by the PSO-optimized hierarchical neural network is in high agreement with the theoretically optimal curve. Based on this dynamically optimized threshold, the system achieved a false negative rate of 0.004% in comprehensive testing covering a signal-to-noise ratio range of 0dB to 20dB, demonstrating that the method provided by this invention can effectively enhance the ability to capture weak signals.

[0118] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamically optimizing the detection threshold of radio frequency signals, characterized in that, The methods include: Acquire the radio frequency signal received by the radio frequency receiving link and the real-time temperature collected synchronously; The radio frequency signal is subjected to time-domain verification and image-domain recognition, and the time-domain verification results and image-domain recognition results are fused to determine the effective pulse signal, the pulse confidence of the effective pulse signal, and the interference intensity. Based on the effective pulse signal, the effective signal and noise are separated and extracted to obtain the effective signal power and noise power, and the instantaneous signal-to-noise ratio is generated by combining the effective signal power and noise power. Based on the pulse confidence, outlier repair is performed on the instantaneous signal-to-noise ratio sequence composed of the instantaneous signal-to-noise ratio, and the real-time signal-to-noise ratio is determined by combining the pre-built dynamic dual-window and attention-based temporal prediction model. Based on the real-time signal-to-noise ratio, interference intensity, real-time temperature, and cumulative working time, a target detection threshold is generated using a hierarchical neural network optimized by the particle swarm optimization algorithm.

2. The method according to claim 1, characterized in that, The radio frequency signal undergoes time-domain verification and image-domain recognition. The time-domain verification results and image-domain recognition results are then fused to determine the effective pulse signal, the pulse confidence level of the effective pulse signal, and the interference intensity, including: The radio frequency signal is subjected to normalized autocorrelation pulse detection, adaptive CFAR noise floor suppression, temperature adaptive Bayesian amplitude verification, and pulse slope verification in sequence to determine the time domain verification confidence probability. The time-domain waveform of the radio frequency signal is mapped into a two-dimensional pixel matrix, and the features of the two-dimensional pixel matrix are extracted by a CNN image recognition model to determine the image recognition confidence probability. The fusion weights of the time-domain verification confidence probability and the image recognition confidence probability are dynamically adjusted based on the signal-to-noise ratio status. The pulse confidence level is obtained by weighting and fusing the time-domain verification confidence probability and the image recognition confidence probability based on the adjusted fusion weights, and the effective pulse signal is determined based on the pulse confidence level. The clutter power is then synchronously quantized to determine the interference intensity.

3. The method according to claim 2, characterized in that, The autocorrelation calculation logic is optimized by introducing a noise variance suppression factor and a multipath delay compensation term. Normalized autocorrelation pulse detection is then performed based on the optimized autocorrelation calculation logic. The multipath delay compensation term is determined based on the multipath delay calculated from the additional distance of multipath propagation of the radio frequency signal. When the multipath interference intensity is greater than or equal to a preset interference threshold, the weight of the multipath delay compensation term is increased. The adaptive CFAR noise suppression includes: estimating the background noise power using a unit average constant false alarm rate algorithm, and calculating a dynamic detection threshold based on the average signal power of the reference unit, the preset false alarm rate, and the number of reference units; The temperature-adaptive Bayesian amplitude verification includes: dynamically adjusting the prior parameters of the Bayesian verification based on the real-time temperature; wherein, the prior parameters include the room temperature prior mean and the temperature correction coefficient, and the temperature correction coefficient is determined based on the temperature difference between the real-time temperature and the room temperature reference. The pulse slope verification includes: calculating the rising edge slope and falling edge slope of the pulse; when the rising edge slope is greater than or equal to a rising edge slope threshold and the falling edge slope is greater than or equal to a falling edge slope threshold, it is determined to be a valid potential pulse.

4. The method according to claim 2, characterized in that, The time-domain waveform of the radio frequency signal is mapped into a two-dimensional pixel matrix. Features are extracted from the two-dimensional pixel matrix using a CNN image recognition model to determine the image recognition confidence probability, including: Using time as the horizontal axis and amplitude as the vertical axis, the time-domain waveform of the radio frequency signal is quantized into grayscale values ​​to generate a two-dimensional pixel matrix; wherein, the horizontal axis covers the sampling points of a preset signal period, and the vertical axis corresponds to the quantization range of the signal power; The CNN image recognition model includes an input layer, at least one convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer receives a two-dimensional pixel matrix, which is then passed through at least one convolutional layer to extract local spatial features. The local spatial features are then subjected to max pooling dimensionality reduction through the pooling layer to obtain image texture features. The image texture features are classified and mapped by the fully connected layer, and the classification probabilities of the valid pulse signal and the false pulse are output by the output layer. The classification probability of the valid pulse signal is used as the confidence probability of image recognition.

5. The method according to claim 1, characterized in that, Based on the effective pulse signal, the effective signal and noise are separated and extracted to obtain the effective signal power and noise power, and the instantaneous signal-to-noise ratio is generated by combining the effective signal power and noise power, including: Centered on the peak position of the pulse output from the normalized autocorrelation pulse detection, the interval corresponding to the width of the main lobe of the pulse is selected as the signal window to extract the effective signal components; The interval outside the signal window is selected as the noise window, and abnormal spikes are removed from the noise signal within the noise window using a preset criterion to obtain a purified noise sequence. Calculate the first time-domain mean square value of the effective signal component within the signal window and the second time-domain mean square value of the purified noise sequence within the noise window, respectively, and transform the first time-domain mean square value and the second time-domain mean square value to obtain the effective signal power and the noise power; The difference between the effective signal power and the noise power is calculated to obtain the instantaneous signal-to-noise ratio.

6. The method according to claim 1, characterized in that, Outlier repair based on impulse confidence in an instantaneous signal-to-noise ratio (SNR) sequence, including: Based on the confidence interval in which the pulse confidence level is located, a differentiated repair strategy is executed on the abnormal signal-to-noise ratio data in the instantaneous signal-to-noise ratio sequence; wherein, the confidence interval includes a high confidence interval, a medium confidence interval, and a low confidence interval; When the pulse confidence level is in the high confidence range, a cubic spline interpolation algorithm is used to repair the abnormal signal-to-noise ratio data; When the pulse confidence level is within the middle confidence level range, a linear prediction algorithm is used to repair the abnormal signal-to-noise ratio data, taking into account the signal-to-noise ratio trend of a preset number of preceding and following pulses. When the pulse confidence level is in the low confidence range, the abnormal signal-to-noise ratio data is removed.

7. The method according to claim 6, characterized in that, By combining a pre-built dynamic dual-window and attention-based temporal prediction model, the real-time signal-to-noise ratio is determined, including: Calculate the variance of the signal-to-noise ratio data within a short-term time window; Based on the comparison results between the variance and the first preset threshold and the second preset threshold, the lengths of the short-term time series window and the long-term time series window are dynamically adjusted. Specifically, when the variance is greater than or equal to the first preset threshold, the short-term time series window is shortened and the long-term time series window is extended; when the variance is less than the second preset threshold, the short-term time series window is extended and the long-term time series window is shortened. Based on the adjusted short-term and long-term time-series windows, the mean short-term signal-to-noise ratio and the mean long-term signal-to-noise ratio are calculated respectively. The short-term signal-to-noise ratio mean, long-term signal-to-noise ratio mean, and variance are input into the attention-based time-series prediction model. The weights of the short-term signal-to-noise ratio mean, long-term signal-to-noise ratio mean, and variance are calculated through the attention mechanism, and the predicted signal-to-noise ratio value for the next time step is output. The real-time signal-to-noise ratio is determined by weighting and fusing the short-term mean signal-to-noise ratio, the long-term mean signal-to-noise ratio, and the predicted signal-to-noise ratio based on adaptive weights.

8. The method according to claim 1, characterized in that, Based on the real-time signal-to-noise ratio, interference intensity, real-time temperature, and cumulative working time, a target detection threshold is generated using a hierarchical neural network optimized by the particle swarm optimization algorithm, including: The hierarchical neural network includes a main network and a compensation network; wherein the network parameters of the main network and the compensation network are determined by optimization using a particle swarm optimization algorithm with the goal of jointly minimizing the false response rate, false alarm rate, and power consumption. The real-time signal-to-noise ratio is input into the main network, the nonlinear relationship between the signal-to-noise ratio and the basic detection threshold is fitted, and the basic detection threshold is output. The real-time signal-to-noise ratio, interference intensity, and real-time temperature are input into the compensation network to determine the environmental compensation amount. Based on the basic detection threshold and the environmental compensation amount, an initial detection threshold is determined; The link aging factor is calculated based on the cumulative working time, and the signal-to-noise ratio range in which the real-time signal-to-noise ratio is located is determined. The initial detection threshold is corrected using the link aging factor and the segmented calibration coefficient corresponding to the signal-to-noise ratio interval to generate the target detection threshold.

9. The method according to claim 8, characterized in that, The signal-to-noise ratio (SNR) range includes a high SNR range, a medium SNR range, and a low SNR range. When the real-time signal-to-noise ratio is in the high signal-to-noise ratio range, the initial detection threshold is adjusted and corrected by using the first segmentation calibration coefficient and the link aging factor to generate the target detection threshold. When the real-time signal-to-noise ratio is in the medium signal-to-noise ratio range, the initial detection threshold is smoothly corrected by the second segmented calibration coefficient and the link aging factor to generate the target detection threshold. When the real-time signal-to-noise ratio is in the low signal-to-noise ratio range, the initial detection threshold is adjusted upward using the third segment calibration coefficient and the link aging factor to generate the target detection threshold.

10. The method according to any one of claims 1 to 9, characterized in that, The method further includes: performing detection control on the radio frequency receiving link based on the target detection threshold, and correcting the parameters of the time domain verification, image domain recognition, outlier repair, and hierarchical neural network based on the state feedback of the detection control.

Images