A method for filtering multipath reflected radio frequency signals in high frequency transmission

Abstract

The application discloses a method for filtering multipath reflection radio frequency signals in high-frequency transmission, comprising the following steps: building a radio frequency intelligent purification processing platform to collect the propagation time delay difference, amplitude attenuation value, frequency offset and phase drift coefficient of the multipath reflection radio frequency signals; inputting the parameters into a multipath signal wave direction intelligent estimation model to output the path wave direction angle value and path weight coefficient; classifying the main path and interference path signals based on the result; calling a lightweight multi-scale convolution filtering algorithm, combining the interference path signal parameter to set the convolution kernel scale parameter, and performing convolution operation on the interference signal to generate a filtering feature matrix; and building a signal cancellation module in the platform to realize interference cancellation through the feature matrix. The method improves the path distinguishing accuracy through dynamic parameter adaptation, guarantees the signal integrity through multi-scale filtering, and improves the filtering efficiency through whole-process cooperation, effectively solves the multipath interference problem in high-frequency transmission, and improves the signal stability.

Description

Technical Field

[0001] This invention relates to the field of fire alarms, fire sensor hardware and algorithms, and particularly to a method for filtering out multipath reflected radio frequency signals in high-frequency transmission. Background Technology

[0002] In high-frequency transmission scenarios, signals are susceptible to multipath reflections due to factors such as building walls, metal obstacles, and atmospheric conditions, resulting in multipath reflected radio frequency (RF) signals with multiple propagation paths. These signals carry complex parameters such as propagation delay differences, amplitude attenuation, frequency shifts, and phase drift, leading to problems like signal superposition interference, increased bit error rate at the receiver, and decreased transmission stability in high-frequency transmission systems. This severely impacts the operational efficiency of technologies reliant on high-frequency transmission, such as communications, radar, and satellite navigation. To address these issues, a signal processing platform adapted to high-frequency scenarios needs to be built, combining path direction estimation and multi-scale filtering techniques to achieve accurate filtering of multipath reflected RF signals. However, current filtering schemes that deeply integrate high-frequency transmission characteristics with multipath signal parameters still lack technological depth, necessitating the development of a systematic filtering method to meet practical application requirements.

[0003] Existing technologies for filtering multipath reflected RF signals have two significant drawbacks: First, existing path direction estimation models often use fixed parameter configurations, failing to adjust the estimation strategy based on dynamic parameters such as real-time propagation delay difference and phase drift coefficient of multipath reflected RF signals in high-frequency transmission. This results in low accuracy in determining the direction of different propagation paths, making it difficult to accurately distinguish between the main path signal and the interference path signal, thus affecting the targeted nature of subsequent filtering operations. Second, existing filtering algorithms often ignore the amplitude attenuation characteristics and frequency offset features of high-frequency signals, using uniform filtering parameters for signal processing. This not only fails to effectively filter interference signals with different attenuation levels and frequency offsets but also easily causes over-filtering of the main path signal, damaging signal integrity. Furthermore, the lack of deep collaboration with RF signal processing platforms makes it difficult to achieve efficient integration of the entire process from parameter acquisition to interference cancellation. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a method for filtering out multipath reflection radio frequency signals in high frequency transmission.

[0005] The technical solution adopted in this invention is a method for filtering multipath reflected radio frequency signals in high-frequency transmission, comprising the following steps: S1, building an intelligent radio frequency purification platform, and collecting an initial parameter set of multipath reflected radio frequency signals in a high-frequency transmission scenario through the platform. The initial parameter set includes the propagation delay difference, signal amplitude attenuation value, carrier frequency offset, and phase drift coefficient of the multipath reflected radio frequency signals; S2, inputting the initial parameter set collected in S1 into an intelligent estimation model of the direction of arrival (DOA) of multipath signals, estimating the direction of propagation of the multipath reflected radio frequency signals through the model, and outputting the DOA angle value and path weight coefficient corresponding to each path; S3, based on the DOA angle value and path weight coefficient output in S2, classifying the multipath reflected radio frequency signals by path and filtering... S4. Select the interference path signals to be filtered out and the main path signals to be retained; S5. Call the lightweight multi-scale convolutional filtering algorithm and set the convolution kernel scale parameters of the algorithm according to the classification results of S3. The convolution kernel scale parameters are correlated with the propagation delay difference and amplitude attenuation value of the multi-path reflected radio frequency signals; S6. Input the interference path signals filtered in S3 into the lightweight multi-scale convolutional filtering algorithm with set parameters. The algorithm performs multi-scale convolution operation on the interference path signals to generate an interference signal filtering feature matrix; S7. Based on the interference signal filtering feature matrix generated in S5, construct a signal cancellation module in the radio frequency intelligent purification processing platform. Through this module, perform interference cancellation operation on the multi-path reflected radio frequency signals in high-frequency transmission and filter out the multi-path reflected radio frequency signals.

[0006] Furthermore, in S2, the intelligent estimation model for the direction of arrival (DOA) of multipath signals calculates the DOA angle value for each path using the following formula: Where, θ i Let be the direction-of-arrival angle of the i-th path, c be the propagation speed of the electromagnetic wave in free space, and τ be the direction of arrival angle. i Let d be the propagation delay difference between the i-th path and the main path, d be the element spacing of the receiving antenna array in the RF intelligent purification processing platform, N be the number of elements in the receiving antenna array, and ω be the propagation delay difference between the i-th path and the main path. i Let be the path weight coefficient of the i-th path.

[0007] Furthermore, in S4, the kernel scale parameter of the lightweight multi-scale convolutional filtering algorithm is set using the following formula: Where, k s Here, A is the kernel scale parameter, α is the amplitude attenuation coefficient, and A is the amplitude attenuation coefficient. i Let f be the amplitude attenuation value of the signal from the i-th interference path, β be the frequency adjustment coefficient, and f be the frequency attenuation value. c This represents the carrier frequency offset of the multipath reflected radio frequency signal.

[0008] Furthermore, in S5, when the lightweight multi-scale convolutional filtering algorithm performs convolution operations on the interference path signal, it uses the following formula to generate the interference signal filtering feature matrix: Among them, M m，n Let x be the value of the element in the m-th row and n-th column of the interference signal filtering feature matrix. m+p，n+q w represents the signal sample value in the (m+p)th row and (n+q)th column of the original data matrix of the interference path signal. p，q Let k be the weight value in the p-th row and q-th column of the convolution kernel in the lightweight multi-scale convolutional filtering algorithm. s is the kernel scale parameter.

[0009] Furthermore, in S6, the signal cancellation module of the radio frequency intelligent purification processing platform cancels interference using the following formula: y(t)=s(t)-γ·M(t), where y(t) is the output signal after filtering out multipath reflected radio frequency signals, s(t) is the original multipath reflected radio frequency signal in high-frequency transmission, γ is the cancellation coefficient, and M(t) is the time-domain interference signal obtained based on the transformation of the interference signal filtering feature matrix.

[0010] Furthermore, in S2, the intelligent estimation model for the direction of arrival of multipath signals corrects the path weight coefficients using the following formula: Where, ω' i ω is the path weight coefficient of the i-th path after correction. i Δφ represents the initial path weight coefficient. i Let max(Δφ) be the phase shift coefficient of the signal along the i-th path. j ) represents the maximum value among all path signal phase drift coefficients.

[0011] Further, S3 includes the following sub-steps: S31, extract the direction-of-arrival angle values ​​of each path output from S2, group the angle values ​​according to the deviation range from the direction-of-arrival angle of the main path, classify the paths with a deviation range greater than a preset threshold as potential interference path groups, and classify the paths with a deviation range less than or equal to the preset threshold as main path association groups; S32, calculate the ratio of the signal amplitude attenuation value of each path in the potential interference path group to the signal amplitude attenuation value of the main path, remove the paths with a ratio greater than a preset ratio threshold from the potential interference path group, and retain the paths with a ratio less than or equal to the preset ratio threshold as interference paths to be screened; S33, obtain the carrier frequency offset of the interference paths to be screened, determine whether the offset exceeds the frequency deviation range allowed by the high-frequency transmission system, mark the interference paths to be screened that exceed the range as interference paths to be filtered out, and re-incorporate the interference paths to be screened that do not exceed the range into the main path association group; S34, integrate the interference paths to be filtered out marked in S33 and the main path association groups divided in S31 to form the final interference path signal set and the main path signal set, thus completing the path classification.

[0012] Further, S4 includes the following sub-steps: S41, extract the propagation delay difference of the interference path signal to be filtered from the classification results of S3, and calculate the average and variance of the propagation delay differences of all interference path signals to be filtered, using the average value as the basic delay parameter of the lightweight multi-scale convolutional filtering algorithm; S42, obtain the amplitude attenuation value of the interference path signal to be filtered, and determine the number of convolutional kernels of the algorithm according to the distribution range of the amplitude attenuation value. Different distribution ranges correspond to different numbers of convolutional kernels, and the larger the amplitude attenuation value, the more convolutional kernels are corresponding to the range; S43, combine the basic delay parameter obtained in S41 and the number of convolutional kernels determined in S42, and initially set the initial scale range of the convolutional kernels. The minimum value of the initial scale range is positively correlated with the basic delay parameter, and the maximum value is positively correlated with the number of convolutional kernels; S44, call the parameter calibration module in the radio frequency intelligent purification processing platform, input the initially set initial scale range of the convolutional kernels into the module, and fine-tune the scale range through the module. The fine-tuning is based on the phase drift coefficient of the multipath reflected radio frequency signal, and finally determine the scale parameter of the convolutional kernel.

[0013] Further, S5 includes the following sub-steps: S51, converting the interference path signals filtered in S3 into a digital signal matrix. The row dimension of this matrix corresponds to the time sampling points of the signal, and the column dimension corresponds to the frequency components of the signal. Each element in the matrix is ​​the signal strength value at the corresponding time sampling point and frequency component; S52, performing convolution operation between the convolution kernel matrix corresponding to the convolution kernel scale parameters determined in S4 and the digital signal matrix obtained in S51. A sliding window method is used during the operation, with the window size consistent with the convolution kernel scale parameters. The sliding step size is set according to the carrier frequency offset of the multipath reflected radio frequency signal; S53, performing feature extraction on the intermediate matrix obtained after the convolution operation, extracting the peak, valley, and mean values ​​of each column in the matrix, and arranging these feature values ​​in the order of time sampling points to form a preliminary feature sequence; S54, performing dimension normalization on the preliminary feature sequence so that the length of the feature sequence is consistent with the number of time sampling points in the digital signal matrix in S51. After normalization, the interference signal filtering feature matrix is ​​obtained.

[0014] A method for filtering multipath reflected radio frequency signals in high-frequency transmission is disclosed. This method is implemented through different units, including: a multipath reflected radio frequency signal parameter acquisition unit, used to acquire the propagation delay difference, signal amplitude attenuation value, carrier frequency offset, and phase drift coefficient of multipath reflected radio frequency signals in a high-frequency transmission scenario; a multipath signal direction-of-arrival intelligent estimation unit, connected to the multipath reflected radio frequency signal parameter acquisition unit, receiving the parameter set output by the unit, estimating the direction of propagation of the multipath reflected radio frequency signals, and outputting the direction-of-arrival angle value and path weight coefficient; and a multipath signal classification unit, connected to the multipath signal direction-of-arrival intelligent estimation unit, receiving the angle value and weight coefficient output by the unit, classifying the multipath reflected radio frequency signals, and... The system outputs interference path signals and main path signals; a lightweight multi-scale convolutional filtering parameter configuration unit, connected to the multi-path signal classification unit, receives the classification results output by the unit and sets the convolution kernel scale parameters of the lightweight multi-scale convolutional filtering algorithm; an interference signal convolution operation unit, connected to both the multi-path signal classification unit and the lightweight multi-scale convolutional filtering parameter configuration unit, receives the interference path signals and convolution kernel scale parameters, performs convolution operations on the interference path signals and outputs the interference signal filtering feature matrix; and a multi-path reflected radio frequency signal interference cancellation unit, connected to the interference signal convolution operation unit, receives the filtering feature matrix output by the unit, constructs a signal cancellation module to perform interference cancellation operations on the multi-path reflected radio frequency signals and outputs the filtered signal.

[0015] Beneficial Effects: This invention proposes a method for filtering multipath reflected radio frequency signals in high-frequency transmission. Addressing the problem that existing path direction estimation models have fixed parameters and cannot adapt to dynamic signal parameters, this method establishes an intelligent radio frequency purification platform to collect parameters such as propagation delay difference and phase drift coefficient of multipath reflected radio frequency signals. These parameters are then input into an intelligent multipath signal direction-of-arrival estimation model for direction estimation and path weight calculation. The estimation strategy can be dynamically adjusted to improve path discrimination accuracy and accurately filter out the main path and interfering path signals, solving the problem of insufficient filtering targeting due to inaccurate direction judgment. Furthermore, it addresses the issue of existing filtering algorithms having uniform parameters that are easily broken. To address the issues of poor signal integrity and incompatibility with the platform, a lightweight multi-scale convolutional filtering algorithm is employed. This algorithm combines the propagation delay difference and amplitude attenuation value of the interference path signal to set appropriate convolutional kernel scale parameters. Multi-scale convolution operations are performed on the interference signal to generate a filtering feature matrix. Then, the platform's signal cancellation module achieves accurate interference cancellation. This approach avoids over-filtering, ensuring the integrity of the main path signal, and achieves end-to-end coordination from parameter acquisition, path estimation, filtering operations to interference cancellation. Ultimately, this improves the signal stability of the high-frequency transmission system, reduces the bit error rate at the receiver, and meets the requirements of communication, radar, and other fields for high-frequency signal transmission quality. Attached Figure Description

[0016] Figure 1This is a flowchart of the method steps of the present invention;

[0017] Figure 2 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, a method for filtering multipath-reflected radio frequency signals in high-frequency transmission includes the following steps:

[0020] S1. Build an intelligent radio frequency purification and processing platform. Through this platform, collect the initial parameter set of multipath reflected radio frequency signals in high frequency transmission scenarios. The initial parameter set includes the propagation delay difference, signal amplitude attenuation value, carrier frequency offset and phase drift coefficient of the multipath reflected radio frequency signals.

[0021] Specifically, S1 is the foundational data acquisition step in the entire multipath reflection RF signal filtering method, providing accurate raw parameter support for subsequent path estimation, signal classification, and filtering. The RF intelligent purification platform, as the hardware carrier for data acquisition, must possess signal reception and parameter extraction capabilities adaptable to high-frequency transmission scenarios. The acquired parameters must cover the key characteristic dimensions of multipath reflection RF signals. The propagation delay difference reflects the time difference in arrival at the receiver from different paths, serving as the core basis for path differentiation. The signal amplitude attenuation value reflects the degree of energy loss during signal propagation, directly affecting the accuracy of subsequent interference path selection. The carrier frequency offset reflects the degree of frequency deviation caused by the propagation environment, serving as an important reference for setting filtering algorithm parameters. The phase drift coefficient reflects the change in signal phase, significantly impacting the accuracy of path direction estimation. If the parameters acquired in this step have errors or are missing, it will directly lead to a decrease in the accuracy of all subsequent processing steps, or even cause filtering failure. Therefore, S1 must ensure the real-time, complete, and accurate acquisition of parameters. In practical implementation, an RF intelligent purification and processing platform is first built, consisting of a high-frequency signal receiving antenna array, a signal conditioning module, a data acquisition card, and a parameter extraction unit. The receiving antenna array is an 8-element uniform linear array, with the element spacing set to 0.5 times the wavelength of the high-frequency signal to ensure effective reception of multipath signals. The signal conditioning module amplifies, filters, and down-converts the received high-frequency signal. The amplification gain is set to 30dB, the filtering bandwidth covers the operating frequency band of the high-frequency transmission system, and the down-converted signal is converted into a 100MHz intermediate frequency signal. The data acquisition card is then activated, with a sampling rate set to 500 MS / s and a sampling duration set to 10 ms. The intermediate frequency signal is sampled and transmitted to the parameter extraction unit. The parameter extraction unit calculates the propagation delay difference using a time-domain analysis algorithm, with an extraction range of 0-100 ns and an accuracy controlled within 1 ns. It calculates the signal amplitude attenuation value using a power detection algorithm, with an extraction range of 0-20 dB and an accuracy controlled within 0.1 dB. It calculates the carrier frequency offset using a frequency estimation algorithm, with an extraction range of -10 kHz to +10 kHz and an accuracy controlled within 1 Hz. It calculates the phase drift coefficient using a phase detection algorithm, with an extraction range of 0-π rad and an accuracy controlled within 0.01π rad. Finally, an initial parameter set containing the above four types of parameters is formed and stored in the platform database.

[0022] S2 inputs the initial parameter set collected in S1 into the intelligent estimation model of multipath signal direction of arrival. The model estimates the direction of propagation of multipath reflected radio frequency signals and outputs the direction of arrival angle value and path weight coefficient corresponding to each path.

[0023] Specifically, S2 is a crucial step in achieving accurate identification of the path direction of multipath reflected radio frequency signals. Through an intelligent multipath signal direction-of-arrival (DOA) estimation model, the discrete parameters collected in S1 are transformed into directional features and weight information suitable for path classification. This intelligent DOA estimation model must be adaptable to dynamic parameters. Unlike traditional fixed models, this model dynamically adjusts its internal calculation logic based on changes in input parameters such as propagation delay difference and phase drift coefficient, ensuring reliable DOA angle values ​​are output under different high-frequency transmission scenarios. The DOA angle value directly reflects the incident direction of each path signal relative to the receiving antenna array and is the core feature distinguishing the main path from interference paths. The path weight coefficient comprehensively reflects the strength and stability of each path signal; a higher weight indicates a greater impact on transmission quality, providing a basis for prioritizing high-weight interference paths. The accuracy of this step directly determines the accuracy of subsequent path classification. Excessive direction estimation error can lead to misclassifying the main path as an interference path or missing key interference paths, thus affecting the filtering effect. In practice, the initial parameter set stored in S1 is first retrieved from the RF intelligent purification processing platform database. The parameters are then validated to remove abnormal data caused by acquisition errors (such as propagation delay differences exceeding the 0-100ns range or negative signal amplitude attenuation values), ensuring that the parameter qualification rate of the input model is not less than 99%. Subsequently, the validated parameters are input into the multipath signal direction-of-arrival intelligent estimation model. The model first normalizes the propagation delay difference, mapping it to a numerical range of 0-1, and then calculates the directional correlation of each path in conjunction with the phase drift coefficient. The correlation calculation window is set to 100 sampling points to balance real-time performance and accuracy. The model iteratively calculates and outputs the direction-of-arrival (DOA) angle values ​​for each path, with the angle range set from -90° to +90° and the calculation error controlled within 0.5°. Simultaneously, it calculates path weight coefficients by combining the signal amplitude attenuation value. The weight coefficient calculation adopts a weighted summation logic, with the weight ratios of propagation delay difference, signal amplitude attenuation value, carrier frequency offset, and phase drift coefficient set to 40%, 30%, 20%, and 10%, respectively. The final output path weight coefficient ranges from 0 to 1, with an accuracy controlled within 0.01. The calculated DOA angle values ​​and path weight coefficients are stored accordingly to form a path feature dataset, which is then transmitted to the subsequent processing module.

[0024] S3, based on the direction-of-arrival angle value and path weight coefficient output by S2, performs path classification on the multipath reflected radio frequency signals, and filters out the interference path signals to be filtered out and the main path signals to be retained.

[0025] Specifically, S3 is the core step in effectively classifying multipath reflected RF signals. Based on the path feature data output by S2, it accurately distinguishes between the main path signal and the interfering path signal, providing a clear target for subsequent targeted filtering. The essence of path classification is to identify interfering paths that negatively impact high-frequency transmission quality through multi-dimensional feature filtering, while retaining the main path that plays a dominant role in transmission. In high-frequency transmission scenarios, the main path typically exhibits characteristics such as stable direction of arrival, small signal amplitude attenuation, and high weighting coefficients, while interfering paths often show characteristics such as a direction of arrival deviating from the main path, unstable signal amplitude attenuation, and large fluctuations in weighting coefficients. This step requires establishing multi-dimensional classification standards to avoid misclassification caused by single-feature judgments—if classification is based solely on the direction of arrival angle, interfering paths with similar directions but large amplitude attenuation may be missed; if classification is based solely on weighting coefficients, low-weight interfering paths with large frequency offsets may be misjudged. Through multi-feature collaborative classification, it is ensured that both the accuracy of main path identification and the coverage of interfering path identification are at a high level, laying the foundation for subsequent filtering processing. In practice, the direction-of-arrival (DOA) angle and path weight coefficient of each path are first extracted from the path feature dataset. A main path reference direction is set (based on the initial main path signal direction acquired by S1, with a deviation allowable range of ±5°). Paths whose DOA angle deviates from the main path reference direction by more than ±5° are marked as "interference paths to be confirmed," and paths with deviations within ±5° are marked as "main paths to be confirmed." Subsequently, the ratio of the signal amplitude attenuation value of the "interference paths to be confirmed" to the signal amplitude attenuation value of the main path is calculated. A ratio threshold of 0.8 is set. If the ratio is greater than 0.8 (i.e., the signal amplitude attenuation of the interference path is close to that of the main path, and it may be a useful signal), it is reclassified as a "main path to be confirmed." If the ratio is less than or equal to 0.8, the "interference path to be confirmed" label is retained. Next, the carrier frequency offset of the "interference path to be confirmed" is extracted. The allowable frequency deviation range for the high-frequency transmission system is set to -5kHz to +5kHz. If the carrier frequency offset exceeds this range, it is ultimately marked as an "interference path to be filtered out". If it is within the range, it is further judged by combining the path weight coefficient. The weight coefficient threshold is set to 0.3. Paths with a weight coefficient less than 0.3 are marked as "interference paths to be filtered out", and those with a weight coefficient greater than or equal to 0.3 are reclassified as "main paths to be confirmed". Finally, the "main paths to be confirmed" are sorted by weight, and the top 3 paths with the highest weight coefficients are selected as the "final main path signals". The remaining signals marked as "interference paths to be filtered out" are integrated into the "final interference path signal set", and the classification is completed and output to S4.

[0026] S4, invoke the lightweight multi-scale convolutional filtering algorithm, and set the convolutional kernel scale parameter of the algorithm according to the classification result of S3. The convolutional kernel scale parameter is correlated with the propagation delay difference and amplitude attenuation value of the multipath reflected radio frequency signal.

[0027] Specifically, S4 is the step that achieves precise adaptation between the lightweight multi-scale convolutional filtering algorithm and the interference path signal. By dynamically setting the convolution kernel scale parameter, the filtering algorithm can efficiently filter out interference signals while avoiding damage to the main path signal, and also ensure the real-time performance of the algorithm. The convolution kernel scale parameter directly determines the filtering algorithm's ability to process interference signals with different characteristics. If the parameter does not match the characteristics of the interference signal, problems such as incomplete filtering or over-filtering will occur. This step establishes a correlation between the convolution kernel scale parameter and the interference path signal parameters, transforming key parameters such as propagation delay difference and signal amplitude attenuation value into the basis for scale setting, breaking the limitations of traditional fixed scale parameters, and enabling the algorithm to dynamically adjust according to changes in the interference signal. At the same time, the lightweight design reduces hardware resource consumption, ensuring that the algorithm can still respond quickly in high-frequency signal high-speed transmission scenarios, providing a guarantee for subsequent real-time filtering. The accuracy of parameter setting in this step directly affects the filtering effect of S5 and the accuracy of interference cancellation in S6. In practice, the propagation delay difference and signal amplitude attenuation value of each interference path are extracted from the final interference path signal set output by S3. The average value of the propagation delay difference and the average value of the signal amplitude attenuation value of all interference paths are calculated as the basic data for setting the scale parameters. The parameter configuration module of the lightweight multi-scale convolutional filtering algorithm is called. The module sets the minimum and maximum values ​​of the convolution kernel scale parameters according to the built-in association rules: the minimum value is positively correlated with the average propagation delay difference; for every 2 nanoseconds increase in the average propagation delay difference, the minimum value increases by 1 unit. The maximum value is positively correlated with the average signal amplitude attenuation value; for every 1 dB increase in the average signal amplitude attenuation value, the maximum value increases by 1 unit. The range is adjusted according to the real-time requirements of the high-frequency transmission system. If the system requires a filtering delay of no more than 1 millisecond, the maximum value is controlled within 20 units; if higher filtering accuracy is required (interference signal suppression ratio greater than 30 dB), the maximum value is increased to 25 units. The set minimum and maximum values ​​are validated to ensure that the difference between them is between 5 and 10 units to avoid redundant calculations. After the validation is passed, the range of scale parameters is determined, and the number of convolution kernels is set to the difference between the maximum and minimum values ​​plus 1, with one convolution kernel corresponding to each scale. The parameter information is stored in the algorithm configuration file to prepare for S5 operation.

[0028] S5 inputs the interference path signal filtered by S3 into a lightweight multi-scale convolutional filtering algorithm with pre-set parameters. The algorithm performs multi-scale convolution operations on the interference path signal to generate an interference signal filtering feature matrix.

[0029] Specifically, S5 is the step of precisely filtering the interference path signal. A lightweight multi-scale convolutional filtering algorithm extracts multi-dimensional features of the interference signal and generates a filtering feature matrix, providing accurate data support for interference cancellation in S6. The algorithm's multi-scale characteristics allow for differentiated processing of interference signals with different frequencies and time delays: small-scale convolutional kernels capture high-frequency details of the interference signal, achieving precise suppression of narrowband interference; large-scale convolutional kernels cover low-frequency wide-delay features of the interference signal, achieving comprehensive suppression of broadband interference. The lightweight design simplifies the computation process and reduces redundant nodes, ensuring that the algorithm can still process in real-time even in high-frequency signal, high-sampling-rate scenarios, avoiding interference signals that cannot be filtered out in time due to computational delays. The filtering feature matrix must completely preserve the time-domain and frequency-domain features of the interference signal. If features are missing or distorted, subsequent interference cancellation will be incomplete, affecting the final filtering effect. Therefore, this step requires strict control over the completeness and accuracy of feature extraction. In practice, time-domain data of the interference path signals are retrieved from the final interference path signal set output by S3. The data length is set to 10,000 sampling points, and the sampling rate is kept consistent with S1 to ensure data synchronization. The time-domain data is converted into a two-dimensional data matrix, with the number of rows corresponding to the number of time periods and the number of columns corresponding to the number of frequency bands, ensuring that each matrix element accurately corresponds to the signal characteristics at a specific time-frequency point. The convolution kernel scale parameters and number are retrieved from the S4 configuration file. The dimension of each convolution kernel matches the column dimension of the data matrix, and the convolution kernel weights are generated according to a specific distribution rule to ensure that the weights are concentrated in the center region of the kernel to reduce the impact on non-target features. The algorithm is launched using a sliding window method for convolution operations. The window size is consistent with the current convolution kernel scale, and the sliding stride is set to 2 sampling points to balance efficiency and accuracy. For each scale of convolution kernel, the operation is performed on each row of the data matrix to generate a filter matrix of the corresponding scale. After the calculation is completed, the weights of the filter matrices of different scales are set according to the frequency distribution of the interference signal (the small scale has a higher weight if the high frequency interference accounts for a higher proportion, and the large scale has a higher weight if the low frequency interference accounts for a higher proportion). Feature fusion is achieved by weighted summation to generate a two-dimensional filter feature matrix. After the matrix is ​​numerically standardized (mapped to the 0-1 range), it is transmitted to the signal cancellation module of the radio frequency intelligent purification processing platform.

[0030] S6, based on the interference signal filtering feature matrix generated by S5, constructs a signal cancellation module in the radio frequency intelligent purification processing platform. This module performs interference cancellation operation on multipath reflected radio frequency signals in high-frequency transmission and filters out multipath reflected radio frequency signals.

[0031] Specifically, S6, through the signal cancellation module of the RF intelligent purification processing platform, achieves precise interference cancellation based on the filter feature matrix generated by S5, completely removing interference components and retaining the useful signal of the main path, directly determining the final signal quality of the high-frequency transmission system. The core of signal cancellation is constructing a cancellation signal with characteristics opposite to the interference signal. This is achieved by superimposing it onto the original signal to suppress interference. This process requires ensuring a high degree of matching between the cancellation signal and the interference signal in amplitude, phase, and frequency; otherwise, residual interference or damage to the main path signal will occur. The module's real-time processing capability can handle the dynamic changes of the interference signal with the propagation environment. By adjusting the cancellation signal parameters in real time, it ensures stable cancellation effects and avoids the lag problem of traditional offline cancellation methods. This step requires strict control of cancellation accuracy. Incomplete cancellation will lead to an increase in the bit error rate at the receiver, failing to meet the high-frequency transmission quality requirements of fields such as communication and radar. In specific implementation, the original multipath reflected RF signal of the high-frequency transmission is retrieved from the signal acquisition module of the RF intelligent purification processing platform. The signal length is set to 10,000 sampling points, and the sampling rate is kept consistent with S1 to ensure synchronization with the cancellation signal. Simultaneously receiving the standardized filter feature matrix transmitted by S5, the signal cancellation module first converts the filter feature matrix into a time-domain cancellation signal through inverse Fourier transform: first, a two-dimensional Fourier transform is performed on the feature matrix to obtain the frequency-domain interference features, and then the carrier frequency and phase information of the original signal are combined to generate a frequency-domain cancellation signal. This is then converted back to a time-domain cancellation signal through inverse Fourier transform, ensuring that the length and sampling rate of the time-domain cancellation signal are completely consistent with the original signal. Amplitude calibration is performed on the time-domain cancellation signal. Based on the signal amplitude attenuation value acquired by S1, the amplitude of the cancellation signal is adjusted to 1.05 times the amplitude of the original interference signal (with a 5% compensation margin to address amplitude errors). Simultaneously, phase calibration is performed. Based on the phase drift coefficient acquired by S1, the phase of the cancellation signal is adjusted to be opposite to the phase of the original interference signal, with the phase deviation controlled within 0.01π radians. The calibrated cancellation signal and the original signal are input to the superposition module, using a real-time synchronous superposition method, with the superposition delay controlled within 10 nanoseconds. After superposition, the signal with filtered interference is output. The output signal is subjected to quality testing. The testing indicators include interference rejection ratio (greater than 30 dB), main path signal amplitude loss (less than 1 dB), and bit error rate (less than 1 × 10^-6). If it fails to meet the test standards, the cancellation signal parameters are readjusted until the test standards are met before outputting to the receiving end.

[0032] Preferably, in S2, the intelligent estimation model for the direction of arrival (DOA) of multipath signals calculates the DOA angle value of each path using the following formula: Where, θ i Let be the direction-of-arrival angle of the i-th path, c be the propagation speed of the electromagnetic wave in free space, and τ be the direction of arrival angle. iLet d be the propagation delay difference between the i-th path and the main path, d be the element spacing of the receiving antenna array in the RF intelligent purification processing platform, N be the number of elements in the receiving antenna array, and ω be the propagation delay difference between the i-th path and the main path. i This is the path weight coefficient for the i-th path (value range: 0-1).

[0033] Specifically, S2 clarifies the method by which the intelligent estimation model for the direction of arrival (DOA) of multipath signals calculates the DOA angle value. This calculation process requires combining key parameters of the multipath reflected RF signals with hardware parameters. In implementation, the propagation speed of electromagnetic waves in free space is first determined, with a fixed value of 300,000 kilometers per second. Then, the element spacing of the receiving antenna array is obtained, typically set to 0.15 meters, and the number of elements is set to 8 based on the requirements of high-frequency transmission scenarios. Subsequently, the propagation delay difference between the i-th path and the main path collected in step S1 is extracted. This value is controlled within the range of 0 to 100 nanoseconds, with an accuracy within 1 nanosecond. Simultaneously, the path weight coefficient of the i-th path initially calculated in step S2 is obtained, with a value between 0 and 1, and an accuracy controlled within 0.01. Through specific computational logic, propagation speed, propagation delay difference, array element spacing, array element quantity, and path weight coefficient are incorporated into the calculation to finally obtain the direction-of-arrival angle value of the i-th path. The unit of this angle value is radians, and the value ranges from -π / 2 to π / 2. The calculation error must be controlled within 0.01 radians to ensure the accuracy of the direction judgment of each path during subsequent path classification and to provide a reliable angle basis for distinguishing between the main path and the interference path.

[0034] Preferably, in S4, the kernel scale parameter of the lightweight multi-scale convolutional filtering algorithm is set by the following formula: Where, k s Let A be the kernel scale parameter, and α be the amplitude attenuation coefficient (range: 0.1-0.9). i Let f be the amplitude attenuation value of the signal from the i-th interference path, β be the frequency adjustment coefficient (range: 0.05-0.5), and f be the amplitude attenuation value of the signal from the i-th interference path. c This represents the carrier frequency offset of the multipath reflected radio frequency signal.

[0035] Specifically, in step S4, the setting method for the convolution kernel scale parameter of the lightweight multi-scale convolutional filtering algorithm is clarified. This setting needs to be closely related to the amplitude and frequency characteristic parameters of the interference path signal. In implementation, first determine the values ​​of the amplitude attenuation coefficient and the frequency adjustment coefficient. The amplitude attenuation coefficient is set to 0.5 based on the attenuation characteristics of high-frequency transmission signals, and the frequency adjustment coefficient is set to 0.2 in combination with the carrier frequency stability requirements. Then, extract the amplitude attenuation value of the i-th interference path signal selected in step S3. This value ranges from 0 to 20 dB, and the accuracy needs to be within 0.1 dB. At the same time, obtain the carrier frequency offset of the multi-path reflected radio frequency signal collected in step S1. Its value ranges from -10 kHz to 10 kHz, and the accuracy is controlled within 1 Hz. Through specific computational logic, the amplitude attenuation coefficient, the amplitude attenuation value of the i-th interference path signal, the frequency adjustment coefficient, and the carrier frequency offset are substituted into the calculation. After the calculation, the result is rounded up to obtain the convolution kernel scale parameter. The unit of this parameter is pixels, and the value ranges from 5 to 25 pixels. This ensures that the convolution kernel scale can adapt to the amplitude attenuation and frequency offset characteristics of the interference path signal, so that the subsequent filtering operation can effectively filter out the interference while avoiding excessive consumption of hardware resources.

[0036] Preferably, in S5, when the lightweight multi-scale convolutional filtering algorithm performs convolution operations on the interference path signal, it uses the following formula to generate the interference signal filtering feature matrix: Among them, M m，n Let x be the value of the element in the m-th row and n-th column of the interference signal filtering feature matrix. m+p，n+q w represents the signal sample value in the (m+p)th row and (n+q)th column of the original data matrix of the interference path signal. p，q Let k be the weight value in the p-th row and q-th column of the convolution kernel in the lightweight multi-scale convolutional filtering algorithm. s is the kernel scale parameter.

[0037] Specifically, section S5 describes the computational process of generating the interference signal filtering feature matrix using the lightweight multi-scale convolutional filtering algorithm. This process relies on the original data of the interference path signal and the convolution kernel parameters. In implementation, first, the convolution kernel scale parameter is determined, set to 10 pixels according to step S4, meaning the number of rows and columns of the convolution kernel are both 10. Then, the original data matrix of the interference path signal is obtained. The row and column dimensions of this matrix are set according to the sampling duration and frequency, typically with 1000 rows and 1000 columns. Each element in the matrix represents the signal sampling value at the corresponding time and frequency point, with the sampling value ranging from -1 to 1 volt, achieving an accuracy within 0.001 volts. Simultaneously, the weight value at each position in the convolution kernel is determined. The weight values ​​are set according to the Gaussian distribution rule, ranging from 0 to 0.5, and the sum of all weight values ​​is 1. By using a sliding window convolution operation, the signal sample values ​​in the corresponding window of the original data matrix are multiplied one by one with the weight values ​​at the corresponding positions of the convolution kernel, and then summed to obtain the element values ​​at the corresponding positions in the interference signal filtering feature matrix. The feature matrix has the same dimension as the original data matrix, and the element values ​​range from 0 to 1, thus completely preserving the time and frequency domain characteristics of the interference signal and providing accurate feature data for subsequent interference cancellation.

[0038] Preferably, in S6, the signal cancellation module of the radio frequency intelligent purification processing platform cancels interference using the following formula: y(t)=s(t)-γ·M(t), where y(t) is the output signal after filtering out multipath reflected radio frequency signals, s(t) is the original multipath reflected radio frequency signal in high-frequency transmission, γ is the cancellation coefficient (value range: 0.8-1.2), and M(t) is the time-domain interference signal obtained based on the transformation of the interference signal filtering feature matrix.

[0039] Specifically, S6 clarifies the computational method for interference cancellation implemented by the signal cancellation module of the radio frequency intelligent purification processing platform. This computation requires combining the original signal and interference characteristic data. During implementation, first, a cancellation coefficient is set. Based on the intensity fluctuation characteristics of the interference signal, its value is set to 1.1 to ensure sufficient interference cancellation while avoiding excessive damage to the main path signal. Then, the original multipath reflected radio frequency signal from the high-frequency transmission is extracted. This signal is a time-domain signal with a sampling duration of 10 milliseconds, a sampling rate of 500 MHz, and an amplitude range between -2 and 2 volts, achieving an accuracy within 0.001 volts. Simultaneously, the interference signal filtering characteristic matrix generated in step S5 is converted into a time-domain interference signal. This signal has the same sampling duration and sampling rate as the original signal, and an amplitude range between 0 and 1.5 volts. By using specific computational logic, the product of the cancellation coefficient and the time-domain interference signal is subtracted from the original multipath reflected radio frequency signal to obtain the output signal after filtering out the multipath reflected radio frequency signal. The amplitude of the output signal is between -1.8 and 1.8 volts, the interference suppression ratio needs to reach more than 30 dB, and the amplitude loss of the main path signal is controlled within 1 dB, thereby achieving effective cancellation of multipath interference and improving the quality of high-frequency transmission signals.

[0040] Preferably, in S2, the intelligent estimation model for the direction of arrival of multipath signals corrects the path weight coefficients using the following formula: Where, ω' i ω is the path weight coefficient of the i-th path after correction (value range: 0-1). i The initial path weight coefficient (value range: 0-1), Δφ i Let max(Δφ) be the phase shift coefficient of the signal along the i-th path. j ) represents the maximum value among all path signal phase drift coefficients.

[0041] Specifically, S2 clarifies the method for correcting the path weight coefficients in the intelligent estimation model of multipath signal direction of arrival. The correction process needs to be combined with phase drift characteristic parameters. In implementation, firstly, the initial path weight coefficient of the i-th path initially calculated in step S2 is extracted. This value is between 0 and 1, with an accuracy controlled within 0.01. Then, the phase drift coefficient of the i-th path signal collected in step S1 is obtained. This coefficient is in radians and ranges from 0 to π, with an accuracy within 0.01 radians. At the same time, the maximum value among the phase drift coefficients of all path signals is calculated. This maximum value is usually between 0.5π and π, depending on the differences in the multipath propagation environment. By using specific computational logic, the initial path weight coefficient of the i-th path is multiplied by the ratio of the phase drift coefficient of the i-th path signal to the maximum value of the phase drift coefficients of all paths, thus obtaining the corrected path weight coefficient of the i-th path. The corrected coefficient remains between 0 and 1 and can more accurately reflect the impact of phase drift on the stability of the path signal, thereby improving the accuracy of the importance judgment of each path during subsequent path classification and avoiding path misjudgment caused by phase drift.

[0042] Preferably, S3 includes the following sub-steps: S31, extract the direction-of-arrival angle values ​​of each path output in S2, group the angle values ​​according to the deviation range from the direction-of-arrival angle of the main path, classify the paths with a deviation range greater than a preset threshold as potential interference path groups, and classify the paths with a deviation range less than or equal to the preset threshold as main path association groups; S32, calculate the ratio of the signal amplitude attenuation value of each path in the potential interference path group to the signal amplitude attenuation value of the main path, remove the paths with a ratio greater than a preset ratio threshold from the potential interference path group, and retain the paths with a ratio less than or equal to the preset ratio threshold as interference paths to be screened; S33, obtain the carrier frequency offset of the interference paths to be screened, determine whether the offset exceeds the frequency deviation range allowed by the high-frequency transmission system, mark the interference paths to be screened that exceed the range as interference paths to be filtered out, and re-incorporate the interference paths to be screened that do not exceed the range into the main path association group; S34, integrate the interference paths to be filtered out marked in S33 and the main path association groups divided in S31 to form the final interference path signal set and the main path signal set, thus completing the path classification.

[0043] Specifically, S3 clarifies the detailed step-by-step implementation process for path classification. Each step requires combining parameters such as direction of arrival (DOA), amplitude attenuation, and frequency offset to achieve accurate classification. During implementation, step S31 first extracts the DOA angle values ​​of each path output from step S2, sets the allowable threshold for the DOA angle deviation of the main path to ±5 degrees, and classifies paths with angle deviations exceeding this threshold as potential interference path groups, and those with deviations within the threshold as main path association groups, thus initially distinguishing path types. Step S32 calculates the ratio of the signal amplitude attenuation value of each path in the potential interference path group to the signal amplitude attenuation value of the main path, sets the ratio threshold to 0.8, removes paths with ratios greater than this threshold from the potential interference path group, and retains paths with ratios less than or equal to the threshold as interference paths to be screened, further filtering through amplitude characteristics. Step S... Step 33: Obtain the carrier frequency offset of the interference path to be screened. Set the allowable frequency deviation range of the high-frequency transmission system to -5 kHz to +5 kHz. Mark the interference path to be screened that has an offset exceeding this range as the interference path to be filtered out. The path that does not exceed the range is re-incorporated into the main path association group. Combine the frequency characteristics to optimize the classification. Step S34: Integrate the interference path to be filtered out marked in step S33 with the main path association group divided in step S31 to form the final interference path signal set and the main path signal set, and complete the classification. The whole process is screened layer by layer through multi-dimensional parameters to ensure that the path classification accuracy is not less than 98%, providing accurate objects for subsequent filtering processing.

[0044] Preferably, S4 includes the following sub-steps: S41, extract the propagation delay difference of the interference path signal to be filtered from the classification results of S3, and calculate the average and variance of the propagation delay differences of all interference path signals to be filtered, using the average value as the basic delay parameter of the lightweight multi-scale convolutional filtering algorithm; S42, obtain the amplitude attenuation value of the interference path signal to be filtered, and determine the number of convolutional kernels of the algorithm according to the distribution range of the amplitude attenuation value. Different distribution ranges correspond to different numbers of convolutional kernels, and the larger the amplitude attenuation value, the more convolutional kernels are corresponding to the range; S43, combine the basic delay parameter obtained in S41 and the number of convolutional kernels determined in S42, and initially set the initial scale range of the convolutional kernels. The minimum value of the initial scale range is positively correlated with the basic delay parameter, and the maximum value is positively correlated with the number of convolutional kernels; S44, call the parameter calibration module in the radio frequency intelligent purification processing platform, input the initially set initial scale range of the convolutional kernels into the module, and fine-tune the scale range through the module. The fine-tuning is based on the phase drift coefficient of the multipath reflected radio frequency signal, and finally determine the scale parameter of the convolutional kernel.

[0045] Specifically, the step-by-step operation of setting the convolution kernel scale parameter in the S4 lightweight multi-scale convolutional filtering algorithm requires adjustment of each step based on the interference path signal parameters and system requirements. During implementation, step S41 extracts the propagation delay difference of the interference path signal to be filtered from the classification results of step S3, calculates the average and variance of the propagation delay differences of all paths to be filtered, sets the average calculation sample size to the total number of paths to be filtered, and controls the variance within 5 nanoseconds. The average is used as the basic delay parameter of the algorithm, providing a time dimension basis for scale setting. Step S42 obtains the amplitude attenuation value of the interference path signal to be filtered, setting the amplitude attenuation value distribution range to 0-5 dB, 5-10 dB, and 10-20 dB, corresponding to 3, 5, and 8 convolution kernels respectively. Larger amplitude attenuation values ​​correspond to more convolution kernels to improve filtering accuracy. Step S43 combines the basic time delay parameters of step S41 and the number of convolution kernels in step S42 to set the initial scale range of the convolution kernels. The initial minimum value is positively correlated with the basic time delay parameters. For every 2 nanoseconds increase in the basic time delay parameters, the minimum value increases by 1 unit. The initial maximum value is positively correlated with the number of convolution kernels. For every 2 increases in the number of kernels, the maximum value increases by 3 units. Step S44 calls the parameter calibration module of the radio frequency intelligent purification processing platform, inputs the initial scale range, and fine-tunes it according to the phase drift coefficient (range 0-π radians) of the multipath reflected radio frequency signals. The fine-tuning amplitude is controlled within ±2 units. Finally, the convolution kernel scale parameters are determined to ensure that the parameters are adapted to the interference characteristics and system performance.

[0046] Preferably, S5 includes the following sub-steps: S51, converting the interference path signal filtered in S3 into a digital signal matrix, where the row dimension of the matrix corresponds to the time sampling points of the signal, the column dimension corresponds to the frequency components of the signal, and each element in the matrix is ​​the signal strength value at the corresponding time sampling point and frequency component; S52, performing convolution operation on the convolution kernel matrix corresponding to the convolution kernel scale parameter determined in S4 and the digital signal matrix obtained in S51, using a sliding window method during the operation, with the window size consistent with the convolution kernel scale parameter, and the sliding step size set according to the carrier frequency offset of the multipath reflected radio frequency signal; S53, extracting features from the intermediate matrix obtained after the convolution operation, extracting the peak, valley, and mean values ​​of each column in the matrix, and arranging these feature values ​​in the order of time sampling points to form a preliminary feature sequence; S54, performing dimension normalization on the preliminary feature sequence so that the length of the feature sequence is consistent with the number of time sampling points in the digital signal matrix in S51, and obtaining the interference signal filtering feature matrix after normalization.

[0047] Specifically, S5 clarifies the step-by-step process of convolution operation and feature matrix generation for interference path signals, ensuring the integrity and accuracy of signal processing at each step. In implementation, step S51 converts the interference path signals filtered in step S3 into a digital signal matrix. The matrix is ​​set with row dimensions corresponding to 10,000 time sampling points (sampling rate 500 MHz) and column dimensions corresponding to 200 frequency components (covering the operating frequency band of high-frequency transmission systems). Matrix elements are the signal strength values ​​at the corresponding time-frequency points (range -1 to 1 volt, precision 0.001 volt), providing structured data for subsequent calculations. Step S52 calls the convolution kernel scale parameters determined in step S4 (e.g., 10 units) and performs a sliding window convolution operation on the corresponding convolution kernel matrix and the digital signal matrix. The window size is set to be consistent with the convolution kernel scale, and the sliding step size is 2 sampling points. The step size setting needs to balance computational efficiency (single sampling point size). Step S53 extracts features from the intermediate matrix after convolution, extracting the peak value (range 0 to 1 volt), valley value (range -1 to 0 volt), and mean value (range -0.5 to 0.5 volt) of each column, and arranges them in the order of time sampling points to form a preliminary feature sequence. The sequence length is consistent with the number of time sampling points. Step S54 performs dimension normalization on the preliminary feature sequence and adjusts the sequence length using linear interpolation to ensure that it is completely consistent with the number of time sampling points of the digital signal matrix in step S51. After normalization, an interference signal filtering feature matrix (dimension 10000×200) is generated, with matrix elements ranging from 0 to 1, providing accurate feature data for interference cancellation in step S6.

[0048] The intelligent direction-of-arrival (DOA) estimation model for multipath signals in this invention is an algorithm that combines dynamic parameters of multipath reflected RF signals to accurately calculate the DOA and weights of each propagation path. It is not a traditional fixed-parameter model; its core lies in improving the accuracy of DOA estimation through dynamic parameter adaptation. The implementation process relies on previously acquired signal parameters: first, it receives the initial parameter set of multipath reflected RF signals transmitted from the RF intelligent purification processing platform, including propagation delay difference, amplitude attenuation value, carrier frequency offset, and phase drift coefficient. The parameters are then validated (abnormal data with excessive errors are removed to ensure a pass rate ≥ 99%). Next, through built-in dynamic calculation logic, the propagation delay difference is mapped to basic DOA-related data. The calculation weights are adjusted by combining the phase drift coefficient (40% weight for propagation delay difference, 10% weight for phase drift coefficient), iteratively outputting the DOA angle values ​​for each path (range -90° to +90°, error ≤ 0.5°) and path weight coefficients (range 0-1, accuracy 0.01). Simultaneously, the weight coefficients can be corrected using the phase drift coefficient to further improve accuracy. The model serves as a core basis for subsequent signal classification, accurately distinguishing between the main path and interference paths, and avoiding directional misjudgments caused by fixed parameters in traditional models. Its significance lies in breaking the limitations of path direction estimation in high-frequency transmission scenarios, enabling the estimation results to match the signal propagation state in real time, laying the foundation for the accuracy of the entire filtering method, and improving the adaptability of high-frequency transmission systems to complex multipath environments.

[0049] The lightweight multi-scale convolutional filtering algorithm in this invention is a signal processing algorithm that dynamically sets filtering parameters and balances filtering accuracy and real-time performance, targeting the interference characteristics of multipath reflected radio frequency signals. Its core lies in the combination of "multi-scale adaptation" and "lightweight design". The implementation involves three steps: The first step is parameter configuration. Based on the multipath signal classification results, the average propagation delay difference and average amplitude attenuation value of the interference path signals are extracted. The range of convolution kernel scale parameters is set (the minimum value is positively correlated with the delay, the maximum value is positively correlated with the attenuation value, and the difference is controlled within 5-10 units). At the same time, the number of convolution kernels is determined (consistent with the difference in scale range) to ensure that the parameters are adapted to the interference characteristics. The second step is convolution operation. The interference path signals are converted into a two-dimensional digital matrix (rows correspond to time sampling points, and columns correspond to frequency components). A sliding window method is used (step size of 2 sampling points). Convolution kernels of different scales (small scale captures high-frequency interference, and large scale covers low-frequency interference) are used to operate on the matrix to generate multiple sets of filter matrices. The third step is feature fusion. The weights of the filter matrices at each scale are set according to the frequency distribution of the interference signals (higher weights for high-frequency interference). The weighted sum is used to generate the interference signal filter feature matrix (the values ​​are normalized to the range of 0-1). This algorithm accurately extracts the features of interference signals, providing data support for subsequent interference cancellation. Simultaneously, its lightweight design (simplifying the computation process and reducing redundant nodes) ensures a computational latency of ≤1ms. Its significance lies in solving the problem of incomplete filtering or signal damage caused by the "one-size-fits-all" approach of traditional filtering algorithms. It enables differentiated processing of interference signals, reduces hardware resource consumption, meets the real-time requirements of high-frequency transmission scenarios, and promotes the development of multipath interference filtering technology towards "precision + efficiency."

[0050] The radio frequency intelligent purification platform in this invention is a hardware and software collaborative system integrating signal acquisition, parameter extraction, path estimation, filtering operations, and interference cancellation. It is not a single hardware device; its core lies in achieving the integration and collaboration of the entire process of filtering multipath reflected radio frequency signals. Its implementation requires the construction of six functional units and ensuring data interoperability: First, a multipath reflected radio frequency signal parameter acquisition unit, consisting of an 8-element uniform linear array antenna (element spacing 0.5 times the signal wavelength), a signal conditioning module (30dB amplification gain, filtering bandwidth covering high-frequency bands), and a data acquisition card (500MS / s sampling rate, 10ms sampling duration), acquiring parameters such as propagation delay difference (0-100ns, accuracy 1ns) and amplitude attenuation value (0-20dB, accuracy 0.1dB); Second, intelligent estimation of multipath signal direction of arrival. The platform comprises six main components: a first, a data acquisition unit; a second, a third, a fourth, a fifth, a sixth, a fifth, a sixth, a seventh, a seventh, a seventh, a eighth, a ninth, a eleventh, and tenth, a eleventh, a eleventh, and a eleventh, a eleventh, a eleventh, and a eleventh, respectively. The first component receives the acquired parameters and runs the estimation model, outputting the direction angle and weight coefficients. The eleventh, a fifth, a sixth, a seventh, a eleventh, a eleventh, and a eleventh, eleventh, eleventh, and eleventh, respectively. The eleventh, a sixth, a seventh, a eleventh, a eleventh, and eleventh, respectively ...

[0051] like Figure 2As shown, a method for filtering multipath reflected radio frequency signals in high-frequency transmission is implemented through different units, including: a multipath reflected radio frequency signal parameter acquisition unit, used to acquire the propagation delay difference, signal amplitude attenuation value, carrier frequency offset, and phase drift coefficient of multipath reflected radio frequency signals in high-frequency transmission scenarios; a multipath signal direction of arrival intelligent estimation unit, connected to the multipath reflected radio frequency signal parameter acquisition unit, receiving the parameter set output by the unit, estimating the direction of propagation of the multipath reflected radio frequency signals, and outputting the direction of arrival angle value and path weight coefficient; and a multipath signal classification unit, connected to the multipath signal direction of arrival intelligent estimation unit, receiving the angle value and weight coefficient output by the unit, and classifying the multipath reflected radio frequency signals into paths. The system classifies and outputs the interference path signal and the main path signal; a lightweight multi-scale convolutional filtering parameter configuration unit, connected to the multi-path signal classification unit, receives the classification result output by the unit and sets the convolution kernel scale parameter of the lightweight multi-scale convolutional filtering algorithm; an interference signal convolution operation unit, connected to both the multi-path signal classification unit and the lightweight multi-scale convolutional filtering parameter configuration unit, receives the interference path signal and the convolution kernel scale parameter, performs convolution operation on the interference path signal and outputs the interference signal filtering feature matrix; and a multi-path reflected radio frequency signal interference cancellation unit, connected to the interference signal convolution operation unit, receives the filtering feature matrix output by the unit, constructs a signal cancellation module to perform interference cancellation operation on the multi-path reflected radio frequency signal and outputs the filtered signal.

[0052] A method for filtering multipath reflected radio frequency (RF) signals in high-frequency transmission is proposed. This method establishes an intelligent RF purification platform, first collecting real-time parameters such as propagation delay difference and phase drift coefficient of the multipath reflected RF signals. These dynamic parameters are then input into an intelligent direction-of-arrival (DOA) estimation model for the multipath signals. The model adjusts its DOA estimation strategy and path weight calculation method based on parameter changes, rather than using fixed parameter configurations. This dynamic adaptation approach accurately identifies the DOA differences of different propagation paths, clearly distinguishing between the main path signal to be retained and the interference path signal to be filtered. This completely solves the problem of inaccurate direction determination in existing technologies, which leads to a lack of targeted filtering operations, laying the foundation for subsequent precise filtering of interference signals.

[0053] This method does not employ uniform filtering parameters. Instead, when calling a lightweight multi-scale convolutional filtering algorithm, it tailors the convolution kernel scale parameters to the characteristics of the interference path signals, such as propagation delay differences and amplitude attenuation values. During the filtering process, the algorithm only performs multi-scale convolution operations on the selected interference path signals, generating a unique interference signal filtering feature matrix. This avoids over-filtering of the main path signal while accurately weakening interference signals with different attenuation levels and frequency shifts, ensuring the integrity of the main path signal. This overcomes the problem of signal distortion or incomplete filtering caused by the "one-size-fits-all" approach of existing filtering algorithms.

[0054] Furthermore, this method enhances the filtering effect and improves the system's practicality through end-to-end technological collaboration, which is one of its key advantages. From parameter acquisition by the RF intelligent purification platform to path analysis by the intelligent estimation model of multipath signal direction of arrival, to interference processing by the lightweight multi-scale convolutional filtering algorithm, and finally to interference cancellation by the platform's signal cancellation module, each link is closely connected and data is interconnected, achieving end-to-end collaboration from "parameter acquisition - path classification - filtering operation - interference cancellation". This collaborative mode not only improves the overall filtering efficiency but also avoids the filtering delay or error caused by independent operation of each module and data fragmentation in existing technologies. Ultimately, it significantly improves the signal stability of high-frequency transmission systems, reduces the bit error rate at the receiver, and is more suitable for practical application needs in fields such as communication and radar.

[0055] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for filtering multipath reflection radio frequency signals in high-frequency transmission, characterized in that, Includes the following steps: S1. Establish an intelligent radio frequency purification platform. This platform collects an initial parameter set of multipath reflected radio frequency signals in high-frequency transmission scenarios. This initial parameter set includes the propagation delay difference, signal amplitude attenuation, carrier frequency offset, and phase drift coefficient of the multipath reflected radio frequency signals. S2. Input the initial parameter set collected in S1 into a multipath signal direction-of-arrival (DOA) intelligent estimation model. The model estimates the propagation path of the multipath reflected radio frequency signals and outputs the DOA angle value and path weight coefficient for each path. S3. Based on the DOA angle value and path weight coefficient output in S2, classify the multipath reflected radio frequency signals, selecting the interference path signals to be filtered out and the main path signals to be retained. S4, Invoke the lightweight multi-scale convolutional filtering algorithm, and set the convolution kernel scale parameter of the algorithm according to the classification result of S3. The convolution kernel scale parameter is correlated with the propagation delay difference and amplitude attenuation value of the multipath reflected radio frequency signal; S5, Input the interference path signal filtered by S3 into the lightweight multi-scale convolutional filtering algorithm with set parameters, and perform multi-scale convolution operation on the interference path signal through the algorithm to generate the interference signal filtering feature matrix; S6, Based on the interference signal filtering feature matrix generated by S5, construct a signal cancellation module in the radio frequency intelligent purification processing platform, and perform interference cancellation operation on the multipath reflected radio frequency signal in high frequency transmission through this module to filter out the multipath reflected radio frequency signal; S3 includes the following sub-steps: S31, extract the direction-of-arrival angle values ​​of each path output in S2, group the angle values ​​according to the deviation range from the direction-of-arrival angle of the main path, classify the paths with a deviation range greater than a preset threshold as potential interference path groups, and classify the paths with a deviation range less than or equal to the preset threshold as main path association groups; S32, calculate the ratio of the signal amplitude attenuation value of each path in the potential interference path group to the signal amplitude attenuation value of the main path, remove the paths with a ratio greater than a preset ratio threshold from the potential interference path group, and retain the paths with a ratio less than or equal to the preset ratio threshold as interference paths to be screened; S33, obtain the carrier frequency offset of the interference paths to be screened, determine whether the offset exceeds the frequency deviation range allowed by the high-frequency transmission system, mark the interference paths to be screened that exceed the range as interference paths to be filtered out, and re-incorporate the interference paths to be screened that do not exceed the range into the main path association group; S34, integrate the interference paths to be filtered out marked in S33 and the main path association groups divided in S31 to form the final interference path signal set and the main path signal set, thus completing the path classification.

2. The method for filtering multipath reflection radio frequency signals in high-frequency transmission according to claim 1, characterized in that, In S2, the intelligent estimation model for the direction of arrival (DOA) of multipath signals calculates the DOA angle values ​​for each path using the following formula: ,in, For the first The direction-of-arrival angle value of the path. The speed at which electromagnetic waves propagate in free space. For the first The difference in propagation delay between the secondary path and the main path. The element spacing of the receiving antenna array in the radio frequency intelligent purification platform The number of elements in the receiving antenna array. For the first The path weight coefficient of each path.

3. The method for filtering multipath reflection radio frequency signals in high-frequency transmission according to claim 1, characterized in that, In S4, the kernel scale parameter of the lightweight multi-scale convolutional filtering algorithm is set by the following formula: ,in, The kernel scale parameter is used for convolution. This is the amplitude attenuation coefficient. For the first The amplitude attenuation value of the signal along the interference path. This is the frequency adjustment factor. This represents the carrier frequency offset of the multipath reflected radio frequency signal.

4. The method for filtering multipath reflection radio frequency signals in high-frequency transmission according to claim 1, characterized in that, In S5, the lightweight multi-scale convolutional filtering algorithm uses the following formula to generate the interference signal filtering feature matrix when performing convolution operations on the interference path signal: ,in, The first element in the interference signal filtering feature matrix Line 1 The element values ​​of the column, The first data in the original data matrix of the interference path signal Line 1 The signal sample values ​​of the column, For the lightweight multi-scale convolutional filtering algorithm, the first convolution kernel is... Line 1 Column weight values, is the kernel scale parameter.

5. A method for filtering multipath reflection radio frequency signals in high-frequency transmission according to claim 1, characterized in that, In S6, the signal cancellation module of the radio frequency intelligent purification processing platform cancels interference using the following formula: ,in, To filter out multipath reflected radio frequency signals from the output signal, This refers to the original multipath reflected radio frequency signal in high-frequency transmission. As the offsetting factor, This is the time-domain interference signal obtained by transforming the feature matrix of the interference signal filter.

6. The method for filtering multipath reflection radio frequency signals in high-frequency transmission according to claim 1, characterized in that, In S2, the intelligent estimation model for the direction of arrival of multipath signals corrects the path weight coefficients using the following formula: ,in, For the revised first The path weight coefficient of each path. These are the initial path weight coefficients. For the first The phase shift coefficient of the path signal, This is the maximum value among all path signal phase drift coefficients.

7. A method for filtering multipath reflection radio frequency signals in high-frequency transmission according to claim 1, characterized in that, S4 includes the following sub-steps: S41, extract the propagation delay difference of the interference path signal to be filtered from the classification results of S3, and calculate the average and variance of the propagation delay difference of all interference path signals to be filtered. Use the average value as the basic delay parameter of the lightweight multi-scale convolutional filtering algorithm; S42, obtain the amplitude attenuation value of the interference path signal to be filtered, and determine the number of convolution kernels of the algorithm according to the distribution range of the amplitude attenuation value. Different distribution ranges correspond to different numbers of convolution kernels, and the larger the amplitude attenuation value, the more convolution kernels are corresponding to the range; S43, combine the basic delay parameter obtained in S41 and the number of convolution kernels determined in S42, and initially set the initial scale range of the convolution kernel. The minimum value of the initial scale range is positively correlated with the basic delay parameter, and the maximum value is positively correlated with the number of convolution kernels; S44, call the parameter calibration module in the radio frequency intelligent purification processing platform, input the initially set initial scale range of the convolution kernel into the module, and fine-tune the scale range through the module. The fine-tuning is based on the phase drift coefficient of the multipath reflected radio frequency signal, and finally determine the scale parameter of the convolution kernel.

8. A method for filtering multipath reflection radio frequency signals in high-frequency transmission according to claim 1, characterized in that, S5 includes the following sub-steps: S51, converting the interference path signals filtered in S3 into a digital signal matrix. The row dimension of this matrix corresponds to the time sampling points of the signal, and the column dimension corresponds to the frequency components of the signal. Each element in the matrix is ​​the signal strength value at the corresponding time sampling point and frequency component; S52, performing convolution operation between the convolution kernel matrix corresponding to the convolution kernel scale parameters determined in S4 and the digital signal matrix obtained in S51. A sliding window method is used during the operation, with the window size consistent with the convolution kernel scale parameters. The sliding step size is set according to the carrier frequency offset of the multipath reflected radio frequency signal; S53, extracting features from the intermediate matrix obtained after the convolution operation, extracting the peak, valley, and mean values ​​of each column in the matrix, and arranging these feature values ​​in the order of time sampling points to form a preliminary feature sequence; S54, performing dimension normalization on the preliminary feature sequence so that the length of the feature sequence is consistent with the number of time sampling points in the digital signal matrix in S51. After normalization, the interference signal filtering feature matrix is ​​obtained.

9. A method for filtering multipath reflection radio frequency signals in high-frequency transmission according to any one of claims 1-8, characterized in that, This method is implemented through different units, including: a multipath reflected radio frequency signal parameter acquisition unit, used to acquire the propagation delay difference, signal amplitude attenuation value, carrier frequency offset, and phase drift coefficient of multipath reflected radio frequency signals in high-frequency transmission scenarios; a multipath signal direction of arrival intelligent estimation unit, connected to the multipath reflected radio frequency signal parameter acquisition unit, receiving the parameter set output by the unit, estimating the direction of propagation of the multipath reflected radio frequency signal, and outputting the direction of arrival angle value and path weight coefficient; and a multipath signal classification unit, connected to the multipath signal direction of arrival intelligent estimation unit, receiving the angle value and weight coefficient output by the unit, classifying the multipath reflected radio frequency signal, and outputting the interference path signal and the main path signal. The system includes: a path signal unit; a lightweight multi-scale convolutional filtering parameter configuration unit connected to the multi-path signal classification unit, which receives the classification results output by the unit and sets the convolution kernel scale parameters of the lightweight multi-scale convolutional filtering algorithm; an interference signal convolution operation unit connected to both the multi-path signal classification unit and the lightweight multi-scale convolutional filtering parameter configuration unit, which receives the interference path signal and the convolution kernel scale parameters, performs convolution operations on the interference path signal, and outputs the interference signal filtering feature matrix; and a multi-path reflected radio frequency signal interference cancellation unit connected to the interference signal convolution operation unit, which receives the filtering feature matrix output by the unit, constructs a signal cancellation module to perform interference cancellation operations on the multi-path reflected radio frequency signal, and outputs the filtered signal.

Images