Ultra-wideband acquisition signal denoising and spurious suppression method based on dual-branch time-frequency fusion

Abstract

The application discloses a kind of based on double branch time-frequency fusion's ultra-wideband acquisition signal denoising and stray suppression method, constructs the processing architecture of parallel cooperation of time domain guide branch and frequency domain denoising branch;Among them, long-range time domain dependence is extracted by anti-aliasing down-sampling, bidirectional time series modeling and time reweighting in time domain guide branch, and compressed time domain characteristics and auxiliary noise prior are output;Residual block cascade based on wavelet transform convolution WTConv is used to extract multi-scale time-frequency features in frequency domain denoising branch, and the decoupling representation of narrow-band sparse stray and wide-band spectral details is realized;Further, the cross-domain guide tensor is constructed using compressed time domain characteristics and auxiliary noise prior, the bounded affine modulation parameter is predicted, and the frequency domain features are dynamically modulated;Then, the modulated frequency domain features and time domain features are fused, and the residual noise is estimated by residual prediction head, and the denoising output is obtained by subtracting from the original noisy signal.

Description

Technical Field

[0001] This invention belongs to the field of optical communication technology, and more specifically, relates to a method for noise reduction and spurious suppression of ultra-wideband acquisition signals based on noise prior-guided dual-branch time-frequency fusion. Background Technology

[0002] With the rapid development of ultra-wideband radar, broadband communication, electronic reconnaissance, high-speed instruments, and high-performance data acquisition equipment, signal acquisition systems are placing higher demands on analog bandwidth, real-time sampling rate, dynamic range, and high-fidelity recovery capabilities. To overcome the sampling rate and bandwidth bottlenecks of single-core analog-to-digital converters, digital bandwidth interleaving (DBI), time interleaving (TI), and their hybrid parallel sampling architectures have become important technical approaches for building ultra-wideband acquisition systems.

[0003] However, in practical ultra-wideband acquisition systems, the raw signal is usually affected by a variety of complex errors and interferences, including broadband random background noise, structured spurs caused by subchannel mismatch, local oscillator leakage, intermodulation distortion, image components, quantization errors, and non-stationary time-varying interference. These errors exhibit coupled distribution characteristics in the time, frequency, and statistical domains, making it difficult for traditional single-domain methods to simultaneously achieve noise reduction, spurious suppression, and high-fidelity signal recovery.

[0004] Conventional wavelet denoising, Kalman filtering, or adaptive filtering algorithms struggle to balance noise suppression with the preservation of useful components when dealing with broadband signals exhibiting strong periodicity and wide spectral distribution. These algorithms often exhibit characteristics similar to low-pass filtering, easily causing a drastic attenuation of the high-frequency amplitude of the useful signal while removing high-frequency noise, severely impacting signal integrity. Traditional convolutional neural networks (CNNs), limited by their receptive field size, struggle to capture long-range temporal dependencies in signals at ultra-high sampling rates; while simple recurrent neural networks (such as RNNs and LSTMs) are not only computationally inefficient when processing extremely long sequences but also lack sensitivity to sparse spurious components in the frequency domain, making accurate identification of weak spurious signals difficult.

[0005] Furthermore, existing cross-domain methods mostly remain at the level of simple feature splicing or static gating, lacking a closed-loop mechanism that uses time-domain priors and noise priors to dynamically adjust the frequency-domain response, thus making it difficult to achieve synergistic suppression of broadband random noise and structured spurious emissions. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for noise reduction and spurious suppression of ultra-wideband acquisition signals based on dual-branch time-frequency fusion. This method addresses the problems in the prior art, such as the difficulty of single-domain models in simultaneously suppressing wideband random noise and narrowband structured spurious signals, the difficulty of achieving cross-domain dynamic modulation through simple splicing, and the easy damage to high-frequency effective components and transient edges during the noise reduction process.

[0007] To achieve the above-mentioned objectives, this invention provides a method for noise reduction and spurious suppression of ultra-wideband acquisition signals based on dual-branch time-frequency fusion, characterized by comprising the following steps:

[0008] (1) Construct a noise prior-guided dual-branch time-frequency fusion network;

[0009] (2) Construct a simulation training dataset and use the simulation training dataset to train the dual-branch time-frequency fusion network until convergence;

[0010] (3) The trained dual-branch time-frequency fusion network is used to process the original noisy signal output by the ultra-wideband acquisition system to obtain the output signal after noise reduction and spurious suppression.

[0011] The objective of this invention is achieved as follows:

[0012] This invention presents a dual-branch time-frequency fusion-based method for denoising and spurious suppression of ultra-wideband acquisition signals. It addresses broadband random noise, structured spurious signals, mismatch errors, and non-stationary interference present in digital bandwidth interleaved acquisition systems and broadband parallel acquisition systems by constructing a parallel and collaborative processing architecture of a time-domain guiding branch and a frequency-domain denoising branch. The time-domain guiding branch extracts long-range time-domain dependencies through anti-aliasing downsampling, bidirectional time-series modeling, and time reweighting, outputting compressed time-domain features and auxiliary noise priors. The frequency-domain denoising branch extracts multi-scale time-frequency features through residual block cascading based on wavelet transform convolution WTConv, achieving decoupling representation of narrowband sparse spurious signals and broadband spectral details. Furthermore, a cross-domain guiding tensor is constructed using compressed time-domain features and auxiliary noise priors to predict bounded affine modulation parameters and dynamically modulate the frequency-domain features. Subsequently, the modulated frequency-domain features are fused with the time-domain features, and residual noise is estimated using a residual prediction head. The denoised output is obtained by subtracting the residual from the original noisy signal.

[0013] Meanwhile, the ultra-wideband acquisition signal noise reduction and spurious suppression method based on dual-branch time-frequency fusion of the present invention also has the following beneficial effects:

[0014] (1) By setting up a time-domain guiding branch and a frequency-domain noise reduction branch, this invention can jointly process broadband random noise and narrowband structured spurious signals in ultra-wideband acquisition signals, thus improving the problem that existing single-domain noise reduction methods cannot simultaneously suppress background noise and spurious signals.

[0015] (2) The present invention utilizes compressed time-domain features and auxiliary noise priors to guide the modulation of frequency domain multi-scale features, which can adaptively adjust the response intensity of different frequency bands according to the changes in noise and spurious, thereby improving the ability to distinguish useful signal components from interference components and enhancing the suppression effect on structured spurious.

[0016] (3) By combining long-sequence dependency modeling in the time domain with multi-resolution feature extraction in the frequency domain, this invention can better preserve the high-frequency lines, phase continuity and transient edge information of the signal while suppressing noise and spurious signals, thereby improving the reconstruction accuracy and waveform fidelity of the output signal.

[0017] (4) This invention is applicable to digital bandwidth interleaved acquisition systems, time interleaved acquisition systems, high-speed digital oscilloscopes, radar receivers, broadband spectrum monitoring equipment and broadband data acquisition equipment, and has strong versatility and engineering application value. Attached Figure Description

[0018] Figure 1 This is a flowchart of the ultra-wideband acquisition signal noise reduction and spurious suppression method based on dual-branch time-frequency fusion of the present invention;

[0019] Figure 2 This is a block diagram of a dual-branch time-frequency fusion network structure;

[0020] Figure 3 This is a block diagram of the time-domain bootstrap branch structure;

[0021] Figure 4 This is a block diagram of the frequency domain noise reduction branch structure;

[0022] Figure 5 This is a block diagram of the core structure of WTRESBlock and WTRESConv;

[0023] Figure 6 This is a block diagram of noise-prior-guided cross-domain modulation and residual recovery structure. Detailed Implementation

[0024] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0025] Example

[0026] In this embodiment, as Figure 1 As shown, the present invention discloses a method for noise reduction and spurious suppression of ultra-wideband acquired signals based on noise prior-guided dual-branch time-frequency fusion, comprising the following steps:

[0027] In this embodiment, the overall architecture of the dual-branch time-frequency fusion network mainly includes the following steps: training dataset construction, dual-branch time-frequency fusion network construction, network training, and signal denoising output.

[0028] (1) Construct a training dataset based on Monte Carlo random sampling;

[0029] Multiple sets of target signals are generated using a Monte Carlo random sampling method. Specifically, within a preset frequency range, amplitude range, and phase range, parameters such as the center frequency, amplitude, initial phase, modulation bandwidth, sweep slope, pulse width, or frequency jump position of the target signal are randomly sampled to generate target signals under different operating conditions. The target signal includes one or more of the following: single-frequency signal, multi-frequency signal, linear frequency modulated signal, step frequency signal, pulse signal, or broadband modulated signal.

[0030] In obtaining the target signal Subsequently, a hybrid interference term is generated based on the physical error model of the ultra-wideband data acquisition system. The mixed interference term is composed of multiple types of hardware non-ideal factors, including broadband background noise, structured spurs, intermodulation distortion components, image components, local oscillator leakage components, subband reconstruction residual components, channel mismatch residual components, quantization noise, and one or more non-stationary time-varying interferences. The noise power, spurious frequency, spurious amplitude, spurious phase, intermodulation component location, image suppression degree, local oscillator leakage intensity, mismatch parameters, and time-varying disturbance parameters of the above interference terms can all be randomly generated using a Monte Carlo method.

[0031] target signal The input is fed into a simulation model that includes subband decomposition, frequency shifting, channel non-ideal response, time-interleaved sampling, quantization noise, and digital reconstruction processes, and a hybrid interference term controlled by Monte Carlo parameters is superimposed to obtain the original noisy simulation signal. ;

[0032] ;

[0033] To improve the coverage of training data across different acquisition conditions, during dataset construction, target signal parameters and system error parameters were independently and randomly generated for each set of samples. This ensured that the samples covered different frequencies, amplitudes, signal-to-noise ratios, spurious densities, intermodulation distortion intensities, and non-stationary interference states. The resulting training samples contained not only broadband random noise but also structured spurious and nonlinear distortion components related to the acquisition system structure and the input signal spectrum. Therefore, the original noisy simulation signal... Preprocessing is performed, including one or more of the following: DC removal, normalization, framing, length alignment, abnormal amplitude limiting, and amplitude normalization; after preprocessing, a training sample set is obtained.

[0034] (2) Construct a noise-prior-guided dual-branch time-frequency fusion network;

[0035] In this embodiment, as Figure 2 As shown, the overall architecture of the dual-branch time-frequency fusion network includes an input preprocessing layer, a time-domain guided branch, a frequency-domain denoising branch, a noise prior-guided cross-domain modulation module, and a fusion residual recovery module.

[0036] The input preprocessing layer is used to perform DC removal and normalization processing on the original noisy signal output by the ultra-wideband acquisition system to obtain the input signal.

[0037] The time-domain guided branch is used to extract long-range time-series dependency information from the input signal and generate compressed time-domain features and auxiliary noise priors;

[0038] like Figure 3 As shown, the time-domain guided branch includes an anti-aliasing downsampling module, a bidirectional long short-term memory network Bi-LSTM, a time-domain reweighting module, a first output head, and a second output head;

[0039] The anti-aliasing downsampling module is used to compress the sequence length of the input signal and suppress high-frequency aliasing artifacts during the downsampling process; the bidirectional long short-term memory network Bi-LSTM is used to extract the forward hidden state and backward hidden state of the input signal respectively, and concatenate the forward hidden state and backward hidden state to obtain the temporal deep features; the temporal reweighting module is used to assign weights to the temporal deep features corresponding to each time step, without compressing the entire time series into a single context vector; the first output head is used to generate compressed temporal features based on the reweighted temporal features, and the second output head is used to generate auxiliary noise priors based on the reweighted temporal features;

[0040] The frequency domain noise reduction branch is used to extract multi-scale frequency domain features from the input signal; such as Figure 4 As shown, the frequency domain noise reduction branch includes an encoder and a processor, both of which are composed of multiple cascaded residual blocks based on wavelet transform convolution WTConv; as Figure 5 As shown in (a), the residual block includes at least a WTConv layer, a normalization layer, a smoothing nonlinear activation layer, a channel mixing layer, and residual connections;

[0041] Among them, such as Figure 5As shown in (b), the WTConv layer includes a local convolution path and a wavelet transform path; the local convolution path is used to extract local neighborhood related features of the input features; the wavelet transform path is used to perform multi-level wavelet decomposition, sub-band feature processing and inverse wavelet reconstruction; the outputs of the local convolution path and the wavelet transform path are additively fused to obtain the WTConv output features;

[0042] The noise prior-guided cross-domain modulation module is used to concatenate the compressed time-domain features and the auxiliary noise prior in the channel dimension to construct a cross-domain guiding tensor; then, the cross-domain guiding tensor and the frequency domain multi-scale features are jointly input into the lightweight mapping module to predict the modulation parameters; subsequently, the modulation parameters are subject to bounded constraints, and the frequency domain multi-scale features are dynamically modulated using the bounded-constrained modulation parameters to obtain the guiding frequency domain features;

[0043] The fusion residual recovery module is used to splice and fuse the compressed time-domain features and the guided frequency-domain features in the channel dimension, and input them into the convolutional fusion and residual prediction sub-network to obtain residual noise estimation; then the residual noise estimation is subtracted from the original noisy signal to obtain the output signal after noise reduction and spurious suppression.

[0044] (3) Train the dual-branch time-frequency fusion network until convergence;

[0045] (3.1) Randomly select a batch of training samples from the training dataset and input each training sample into the dual-branch time-frequency fusion network;

[0046] (3.2) Extract compressed temporal features and auxiliary noise priors through temporal-domain guided branching;

[0047] The input signal is first processed by an anti-aliasing downsampling module to compress the length of the input sequence, reducing sequence redundancy and suppressing high-frequency aliasing artifacts introduced by direct downsampling, resulting in a low-resolution temporal feature sequence. This low-resolution temporal feature sequence is then input into a bidirectional long short-term memory (Bi-LSTM) network to obtain the forward hidden state. With backward hidden state Then, the two are spliced ​​together to form a temporal deep feature. .

[0048] Furthermore, attention scores are calculated for the temporal deep features corresponding to each time step, and the temporal attention weights are obtained after normalization. The attention scores... and normalized weights It can be represented as:

[0049] ;

[0050] ;

[0051] in, , , and For learnable parameters, These are the normalized weights for the corresponding time steps.

[0052] The temporal-guided branch does not compress all temporal features into a single context vector. Instead, it reweights the features at each time step element-wise to maintain temporal resolution and highlight the feature responses of key time segments. The reweighted temporal features can be represented as:

[0053] ;

[0054] After obtaining the reweighted temporal features, the temporal guidance branch outputs two types of prior information through a dual-head output structure: first, compressed temporal features obtained through the feature compression head. First, it is used for subsequent cross-domain modulation and feature fusion; second, it obtains auxiliary noise priors through the noise prediction head. It is used to describe the time-domain distribution characteristics of noise and spurious signals in the input signal.

[0055] In this embodiment, compressed time-domain features With auxiliary noise prior After upsampling, the time length is restored to the time length corresponding to the original input signal, so as to align with the frequency domain features in the time dimension.

[0056] (3.3) Extracting multi-scale features in the frequency domain through a frequency domain denoising branch;

[0057] The input signal is input into an encoder and processor consisting of multiple cascaded WTResBlocks. Local spectral variation features are extracted using the local convolution path in the WTConv layer, multi-resolution sub-band features are extracted using the wavelet transform path, and feature transfer is enhanced through residual connections to obtain multi-scale features in the frequency domain.

[0058] The wavelet transform path can be represented as:

[0059] ;

[0060] ;

[0061] ;

[0062] in, Indicates the first Subband features after wavelet decomposition This represents a learnable mapping that processes subband features. This represents discrete wavelet decomposition. This indicates inverse wavelet reconstruction.

[0063] (3.4) Perform noise prior-guided cross-domain modulation;

[0064] like Figure 6 As shown, after obtaining the compressed time-domain features Auxiliary noise prior After considering the multi-scale features in the frequency domain, noise-prior-guided cross-domain modulation is performed.

[0065] Specifically, compressing time-domain features With auxiliary noise prior Construct a cross-domain guiding tensor by splicing along the channel dimension. :

[0066] ;

[0067] The cross-domain guiding tensor and frequency-domain multi-scale features are jointly input into a lightweight mapping module to predict modulation parameters. and ; for the modulation parameters and By applying hyperbolic tangent function constraints, bounded modulation parameters are obtained. and ;

[0068] ;

[0069] ;

[0070] in, and This is a preset coefficient used to control the modulation amplitude range;

[0071] use and Dynamically modulate the multi-scale features in the frequency domain to obtain the guided frequency domain features;

[0072] ;

[0073] in, Represents the original frequency domain characteristics. This represents the modulated pilot frequency domain characteristics. This represents element-wise multiplication.

[0074] (3.5) Perform feature fusion and residual prediction;

[0075] After completing the cross-domain modulation, the time-domain features will be compressed. With modulated pilot frequency domain characteristics The data is stitched together along the channel dimension to form a fused feature:

[0076] ;

[0077] The fusion features are input into the fusion prediction subnetwork, which outputs a residual noise estimate. .

[0078] The residual noise estimation is used to represent the interference components to be removed from the original input signal, including but not limited to broadband random noise, structured spurious noise, and other non-ideal interference components.

[0079] Finally, the denoised output signal is obtained according to the following formula:

[0080]

[0081] in, The original noisy input signal, For residual noise estimation, This is the noise-reduced output signal.

[0082] Since the object estimated by the fusion prediction subnetwork is the residual interference component rather than the complete target signal, the network can focus more on modeling noise and spurious signals, thereby improving denoising and spurious signal suppression performance.

[0083] (3.6) Calculate the total loss function value after this round of training;

[0084] In this embodiment, it is assumed that the current training batch contains There are training samples, and the sequence length of each training sample is . For the first There are training samples, and the original noisy input signal is denoted as . The target clean signal is recorded as The denoised signal output by the dual-branch time-frequency fusion network is denoted as The auxiliary noise prior of the noise prior branch output is denoted as Based on the original noisy input signal Clean signal with target It can construct realistic noise components. ,Right now:

[0085] ;

[0086] in, Indicates the first The training samples contain a real residual noise component composed of one or more of the following: broadband random noise, structured spurious noise, intermodulation distortion, mirror component, local oscillator leakage component, subband reconstruction residual component, and non-stationary time-varying interference.

[0087] To simultaneously constrain the temporal waveform consistency, auxiliary noise prior learning capability, and frequency domain spectral fidelity of the denoised output, this embodiment constructs a total loss function. :

[0088] ;

[0089] in, For temporal reconstruction loss; To assist in predicting loss based on noise; This represents frequency domain consistency loss. , and These are weighting coefficients used to adjust the contribution of each loss term to the total loss function.

[0090] The temporal reconstruction loss described in this embodiment Used to constrain the noise-reducing output signal Clean signal with target The formula for calculating the difference in the time-domain waveform is as follows:

[0091] ;

[0092] In the formula, Indicates the first The denoised output signal of the training sample at the th training sample... The amplitude at each sampling point This indicates that the corresponding target clean signal is at the 1st The amplitude at each sampling point. By minimizing This allows the denoised output signal to approximate the target clean signal in terms of both time-domain amplitude and overall waveform.

[0093] The auxiliary noise prediction loss described in this embodiment Used to constrain auxiliary noise priors Compared with the true noise component The consistency between them is calculated using the following formula:

[0094] ;

[0095] In the formula, The auxiliary noise prior, representing the output of the noise prior branch, is in the first... The amplitude at each sampling point Indicates by The constructed true noise component is in the first... The amplitude at each sampling point. By minimizing This allows the noise prior branch to learn the distribution characteristics of noise and spurious signals in the input signal, thereby improving the network's ability to identify broadband random noise, structured spurious signals, and non-stationary interference.

[0096] The frequency domain consistency loss described in this embodiment Used to constrain the noise-reducing output signal Clean signal with target Consistency in the amplitude spectrum in the frequency domain. Let... Represents the Fast Fourier Transform. This represents the number of frequency points involved in the loss calculation. The calculation formula is:

[0097] ;

[0098] In the formula, Indicates the first The denoised output signal of the training sample at the th training sample... The amplitude spectrum at each frequency point This indicates that the corresponding target clean signal is at the 1st The amplitude spectrum at each frequency point. By minimizing This can reduce local stray residues and damage to effective spectral components.

[0099] In one implementation, the weighting coefficient , and It can be a fixed value or dynamically adjusted based on the training rounds. For example, increasing it at the beginning of training. The value of is chosen so that the network learns the prior distribution of noise and spurious signals first; the value is reduced in the later stages of training. The value of is taken, and increased. or The relative effect of these factors enables the network to further improve the time-domain waveform recovery capability and frequency-domain spectral fidelity of the denoised output signal.

[0100] Through the aforementioned total loss function, this embodiment can simultaneously constrain time-domain waveform recovery, auxiliary noise prior learning, and frequency-domain spectral shape preservation during training. Among these, Used to improve the consistency between the denoised output and the target signal in the time domain. Used to improve the network's ability to identify noise and spurious components. This is used to preserve the frequency domain spectral shape of the denoised output. Training a dual-branch time-frequency fusion network based on this total loss function is beneficial for maintaining the time-domain waveform and frequency-domain structure of the target signal while suppressing broadband random noise and structured spurious signals.

[0101] (3.7) Perform backpropagation and parameter update based on the total loss value;

[0102] Repeat steps (3.1) to (3.6) until the bi-branch time-frequency fusion network converges, and the trained bi-branch time-frequency fusion network is obtained.

[0103] (4) The original noisy signal output by the ultra-wideband acquisition system is acquired in real time, and then the trained dual-branch time-frequency fusion network is used to process the original noisy signal output by the ultra-wideband acquisition system to obtain the output signal after noise reduction and spurious suppression.

[0104] Through the above implementation methods, this invention can jointly model and suppress broadband random noise, structured spurious signals, intermodulation distortion, image components, local oscillator leakage components, and non-stationary time-varying interference in ultra-wideband acquisition systems. Compared with methods that rely solely on single time-domain or frequency-domain processing, this invention extracts long-range time-series dependencies and noise prior information through a time-domain guided branch, extracts multi-scale frequency-domain structural features through a frequency-domain denoising branch, and achieves time-frequency feature fusion through cross-domain modulation and residual recovery. This allows it to suppress mixed noise and spurious components while preserving the time-domain waveform and frequency-domain structure of the target signal.

[0105] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for denoising and spurious suppression of ultra-wideband acquisition signals based on dual-branch time-frequency fusion, characterized in that, Includes the following steps: (1) Construct a noise prior-guided dual-branch time-frequency fusion network; (2) Construct a simulation training dataset and use the simulation training dataset to train the dual-branch time-frequency fusion network until convergence; (3) The trained dual-branch time-frequency fusion network is used to process the original noisy signal output by the ultra-wideband acquisition system to obtain the output signal after noise reduction and spurious suppression.

2. The method for noise reduction and spurious suppression of ultra-wideband acquisition signals based on dual-branch time-frequency fusion according to claim 1, characterized in that, The overall architecture of the dual-branch time-frequency fusion network includes an input preprocessing layer, a time-domain guided branch, a frequency-domain noise reduction branch, a noise prior-guided cross-domain modulation module, and a fusion residual recovery module. The input preprocessing layer is used to perform DC removal and normalization processing on the original noisy signal output by the ultra-wideband acquisition system to obtain the input signal. The time-domain guiding branch is used to extract long-range temporal dependency information from the input signal and generate compressed time-domain features and auxiliary noise priors; the time-domain guiding branch includes an anti-aliasing downsampling module, a bidirectional long short-term memory network Bi-LSTM, a time-domain reweighting module, a first output head and a second output head; The anti-aliasing downsampling module is used to compress the sequence length of the input signal and suppress high-frequency aliasing artifacts during the downsampling process; the bidirectional long short-term memory network Bi-LSTM is used to extract the forward hidden state and backward hidden state of the input signal respectively, and concatenate the forward hidden state and backward hidden state to obtain the temporal deep features; the temporal reweighting module is used to assign weights to the temporal deep features corresponding to each time step, without compressing the entire time series into a single context vector; the first output head is used to generate compressed temporal features based on the reweighted temporal features, and the second output head is used to generate auxiliary noise priors based on the reweighted temporal features; The frequency domain denoising branch is used to extract frequency domain multi-scale features from the input signal; the frequency domain denoising branch includes an encoder and a processor, both of which are composed of multiple cascaded residual blocks based on wavelet transform convolution WTConv; the residual block includes a WTConv layer, a normalization layer, a smoothing nonlinear activation layer, a channel mixing layer, and residual connections. The WTConv layer includes a local convolutional path and a wavelet transform path. The local convolutional path is used to extract local neighborhood-related features of the input features. The wavelet transform path is used to perform multi-level wavelet decomposition, sub-band feature processing, and inverse wavelet reconstruction. The outputs of the local convolutional path and the wavelet transform path are additively fused to obtain the WTConv output features. The noise prior-guided cross-domain modulation module is used to concatenate the compressed time-domain features and the auxiliary noise prior in the channel dimension to construct a cross-domain guiding tensor; then, the cross-domain guiding tensor and the frequency domain multi-scale features are jointly input into the lightweight mapping module to predict the modulation parameters; subsequently, the modulation parameters are subject to bounded constraints, and the frequency domain multi-scale features are dynamically modulated using the bounded-constrained modulation parameters to obtain the guiding frequency domain features; The fusion residual recovery module is used to splice and fuse the compressed time-domain features and the guided frequency-domain features in the channel dimension, and input them into the convolutional fusion and residual prediction sub-network to obtain residual noise estimation; then the residual noise estimation is subtracted from the original noisy signal to obtain the output signal after noise reduction and spurious suppression.

3. The method for noise reduction and spurious suppression of ultra-wideband acquisition signals based on dual-branch time-frequency fusion according to claim 1, characterized in that, The training process of the dual-branch time-frequency fusion network is as follows: (3.1) Construct the training dataset; Multiple sets of target signals are generated using the Monte Carlo random sampling method. ; In obtaining the target signal Subsequently, based on the noise, spurious, and non-ideal error characteristics of the ultra-wideband acquisition system, a hybrid interference term is generated based on the physical error model of the ultra-wideband data acquisition system. , Indicates time; Each set of target signals is superimposed with the mixed interference term to obtain the original noisy simulation signal. ; ； For the original noisy simulation signal The training sample sequence is obtained by performing DC removal, normalization, framing, and amplitude standardization. (3.2) Randomly select a batch of training samples from the training dataset and input each training sample into the dual-branch time-frequency fusion network; (3.3) Extract compressed temporal features and auxiliary noise priors through temporal-domain guided branching; The input signal is first fed into an anti-aliasing downsampling module to obtain a low-resolution temporal feature sequence. This low-resolution temporal feature sequence is then fed into a bidirectional long short-term memory (Bi-LSTM) network to obtain forward and backward hidden states. The forward and backward hidden states are concatenated to obtain deep temporal features. Then, the deep temporal features at each time step are reweighted based on an attention scoring function to obtain reweighted temporal features. Finally, the reweighted temporal features are fed into the first and second output heads respectively to obtain compressed temporal features and auxiliary noise priors. (3.4) Extracting multi-scale features in the frequency domain through a frequency domain denoising branch; The input signal is input into an encoder and processor consisting of multiple cascaded WTResBlocks. Local spectral variation features are extracted using the local convolution path in the WTConv layer, multi-resolution sub-band features are extracted using the wavelet transform path, and feature transfer is enhanced through residual connections to obtain multi-scale features in the frequency domain. (3.5) Perform noise prior-guided cross-domain modulation; By concatenating the compressed time-domain features and auxiliary noise priors, a cross-domain guiding tensor is obtained. The cross-domain guiding tensor and frequency domain multi-scale features are jointly input into a lightweight mapping module to predict modulation parameters. and ; for the modulation parameters and By applying hyperbolic tangent function constraints, bounded modulation parameters are obtained. and ;use and Dynamically modulate the multi-scale features in the frequency domain to obtain the guided frequency domain features; The calculation formula for the dynamic modulation is as follows: ； in, Represents the original frequency domain characteristics. This represents the modulated pilot frequency domain characteristics. Represents element-wise multiplication; (3.6) Perform feature fusion and residual prediction; The compressed time-domain features and the guided frequency-domain features are concatenated and fused along the channel dimension to obtain the fused features; the fused features are then input into the convolutional fusion and residual prediction subnetwork to output residual noise estimation; the denoised output signal is obtained according to the following formula: ； In the formula, ŝ represents the residual noise estimation, and ŝ represents the output signal after noise reduction and spurious suppression. (3.7) Calculate the total loss value during the training process; ； in, This is the temporal reconstruction loss, used to constrain the difference between the denoised output and the target signal; The auxiliary noise prediction loss is used to constrain the consistency between the auxiliary noise prior and the actual noise components. Frequency domain consistency loss is used to constrain the spectral fidelity of the denoised output in the frequency domain; , and These are the weighting coefficients; (3.8) Perform backpropagation and parameter update based on the total loss value; Repeat steps (3.2) to (3.7) until the bi-branch time-frequency fusion network converges, and the trained bi-branch time-frequency fusion network is obtained.

Images