Broadband unknown interference real-time suppression method based on spectral domain self-supervision

Abstract

The invention discloses a spectral domain self-supervision-based broadband unknown interference real-time suppression method without manual labeling, which comprises the following steps of: firstly, performing spectral domain analysis on a baseband IQ signal of a receiving end, and randomly generating a frequency band mask to construct a double-view sample; then, performing joint modeling on the mask spectrum and the original spectrum by using a self-supervised contrast learning model to obtain deep representation capable of distinguishing interference and signal components; thirdly, introducing a frequency gating attention network on the basis of a frozen encoder, adaptively outputting an amplitude-phase dual-path suppression mask according to encoding characteristics, and completing spectral domain filtering of pilot frequency trailing and broadband sweep frequency interference; and finally, performing inverse Fourier transform on the power spectrum after interference suppression to obtain a time domain signal. Through the self-supervised representation learning and lightweight attention technology, the precision and real-time performance of broadband interference suppression are effectively improved, and meanwhile, the self-adaptive capability of the system in a changeable wireless environment is enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication signal processing and deep learning, and particularly relates to a wideband unknown interference real-time suppression method based on spectral domain self-supervised learning without manual annotation. BACKGROUND

[0002] With the rapid popularization of 5G / 6G mobile communication, satellite Internet and Internet of Things, the air spectrum is increasingly crowded, and wideband and variable component unknown interference (such as sweep, frequency hopping and phase modulation tail interference) brings significant challenges to the signal quality of the receiver. Traditional interference suppression methods mainly rely on hardware notch filters, auxiliary radio frequency reference paths or pilot-based adaptive FIR cancellation. These schemes often require prior calibration or injection of known signals, and it is difficult to maintain suppression accuracy in a rapidly changing wireless environment, and additional requirements are put forward for the power consumption and volume of the terminal device.

[0003] In recent years, deep learning methods have been introduced into the communication physical layer for tasks such as interference identification and beamforming. However, most existing models use supervised learning strategies, which require a large amount of training data with accurate interference labels, and are prone to failure when new or variant interference occurs. On the other hand, self-supervised contrastive learning has shown strong representation capabilities in image and speech fields, but its application in wideband interference suppression scenarios is still in its infancy: how to construct an effective spectral domain self-supervised task, separate interference and useful signals under no-label conditions, and perform real-time inference on edge devices and dynamically adapt to channel conditions, are still unresolved problems.

[0004] Therefore, there is an urgent need for a new method of wideband unknown interference suppression that can break free from manual annotation and pilot dependence, while having lightweight real-time inference capability, to improve the reliability and adaptability of communication systems in complex wireless environments. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a wideband unknown interference real-time suppression method based on spectral domain self-supervision without manual annotation. Under the premise of fixed sampling rate, fixed mask strategy and fixed network structure, the method can realize real-time suppression of 5MHz-20MHz wideband unknown interference by combining spectral domain contrastive learning and frequency-gated attention.

[0006] The purpose of the present application is achieved by the following technical solution: a wideband unknown interference real-time suppression method based on spectral domain self-supervision, comprising the following steps:

[0007] S1, signal acquisition and spectral domain preprocessing; performing fast Fourier transform on the received baseband IQ signal to obtain a power spectrum; simultaneously performing baseline correction and amplitude limiting on the spectral data to obtain preprocessed spectrum;

[0008] S2, random mask view construction: randomly select non-overlapping frequency bands in the pre-processed spectrum domain and perform amplitude masking to generate a mask spectrum; and the mask spectrum and the original pre-processed spectrum are combined to form a double-view sample;

[0009] S3, self-supervised contrastive learning representation training: inputting the double-view sample into a feature encoder, training the encoder through a contrastive learning loss function, and freezing the encoder parameters after training;

[0010] S4, frequency-gated interference suppression: introducing a frequency-gated attention network, dynamically generating an amplitude-phase dual-channel suppression mask based on the encoder output feature, and multiplying the mask with the pre-processed spectrum to obtain a power spectrum after interference suppression;

[0011] S5, time domain reconstruction and output: performing inverse Fourier transform on the power spectrum after interference suppression to obtain the recovered time domain signal.

[0012] The beneficial effects of the present application are: the present application closely combines the spectrum domain self-supervised contrastive learning and the frequency-gated attention mechanism, realizes the real-time suppression of wideband unknown interference without relying on pilot signals and artificial labels: the self-supervised encoder can automatically distinguish useful signal and interference features, and the amplitude-phase dual-channel gated mask can suppress interference while maintaining signal intrinsic coherence; the online incremental update mechanism can adaptively adjust the weight when the channel condition changes or new type of interference appears, and always maintain stable suppression ability; the overall algorithm is lightweight and easy to quantize and deploy, which can be quickly migrated to various hardware platforms such as satellites, base stations and Internet of Things terminals, greatly improving the reliability, resource utilization efficiency and operation convenience of wireless communication systems. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is the overall flowchart of the wideband unknown interference suppression method based on spectrum domain self-supervision of the present application;

[0014] Figure 2 is the internal flowchart of the random mask view construction of the present application;

[0015] Figure 3 is the internal flowchart of the self-supervised contrastive learning representation training of the present application;

[0016] Figure 4 is the internal flowchart of the frequency-gated interference suppression of the present application. DETAILED DESCRIPTION

[0017] To solve the technical problems that the existing interference suppression scheme needs to rely on pilots or a large number of labels and is difficult to adapt to wideband unknown interference in real time, the present application provides a wideband unknown interference real-time suppression method based on spectral domain self-supervision without manual annotation. The method realizes real-time suppression of 5MHz-20MHz wideband unknown interference by combining spectral domain contrast learning and frequency-gated attention under the premise of fixed sampling rate, fixed mask strategy and fixed network structure.

[0018] The technical solutions of the present application will be further described below with reference to the drawings.

[0019] As shown in the figure, a wideband unknown interference real-time suppression method based on spectral domain self-supervision without manual annotation of the present application comprises the following steps: Figure 1

[0020] S1, signal acquisition and spectral domain preprocessing; the received baseband IQ signal is subjected to fast Fourier transform to obtain a power spectrum; at the same time, baseline correction and amplitude limitation are performed on the spectral data to obtain preprocessed spectrum;

[0021] The specific implementation method is: receiving a baseband IQ signal x(t), and performing discrete sampling on the received baseband complex signal x(t) at a fixed sampling rate Fs, wherein t is a time sequence, and the length of each frame sample after sampling is L;

[0022] In this embodiment, the receiver digitizes the baseband signal after down-conversion of the 2.4GHz carrier at a fixed sampling rate of 40MSps, and each sampling point is generated by a 12-bit ADC, which is represented as x[n] = x I [n] + jx Q [n], n is a discrete point sequence, j is an imaginary unit, x I [n] represents the in-phase component of the signal, and x Q [n] represents the quadrature component of the signal. The acquisition ring buffer is set to 65536 points; then, continuous frames are extracted from the buffer with a frame length of L = 4096 points and an overlap of 2048 points (i.e. 50%).

[0023] To suppress sidelobe leakage, a window function w[n] is applied to each frame of length L sample to obtain the windowed sequence x w [n]. And 4096 fast Fourier transform is performed on x w [n] to obtain complex spectrum X[k]:

[0024]

[0025] wherein k is a discrete frequency index; w[n] is a Hann window,

[0026] ​After obtaining the power spectrum X[k], continue to calculate the power P[k] = |X[k]| 2 , and estimate the background noise baseline P b [k] = med(P[k-1], P[k], P[k+1]) to obtain the baseline corrected spectrum P c [k] = P[k] - P b [k];

[0027] To suppress occasional spikes, first calculate the root mean square amplitude of the corrected spectrum Then calculate the threshold according to the fixed clipping coefficient α ∈ (0, 1):

[0028]

[0029] And perform amplitude limiting:

[0030] X lim [k] = min(|X[k]|, A max )e j∠X[k]

[0031] If the amplitude of a certain frequency component exceeds A max , replace the original amplitude with A max and keep the phase; finally form the preprocessed spectrum X in [k] = X lim [k] for use in the next step.

[0032] S2, random mask view construction: randomly select non-overlapping frequency bands in the preprocessed spectrum domain and perform amplitude masking to generate a mask spectrum; the mask spectrum and the original preprocessed spectrum form a double-view sample; the flow is shown in Figure 2 , which specifically includes:

[0033] S2.1, random frequency band selection: in the 4096-point preprocessed spectrum X in [k] obtained in the last step, first set a positive integer K = 4 to represent the number of frequency bands to be shielded; divide the non-negative frequency half-axis {k | 0 ≤ k < 2047} (positive frequency half-axis) into several candidate intervals, each with a width of P = 32 frequencies; then use discrete uniform distribution to generate a starting index s i (i = 1, …, 4) for each mask frequency band (the frequency band to be shielded); then set a constant bandwidth P = 32, and construct an index set Ω i = {s i , s i +1, …, s i +31} for each segment; if a certain Ω i has an intersection with an existing Ω j , j < i, then resample si , until all Ω i are mutually exclusive; finally, Ω is regarded as the mask index set of the current frame signal; in this way, it can be ensured that the shielded areas are sufficiently spaced apart and do not overlap with each other.

[0034] S2.2, Mask spectrum generation: define a binary mask function on the index set Ω:

[0035]

[0036] Use this function to obtain the mask spectrum:

[0037] X mask [k] = M[k] · X in [k]

[0038] And the original phase ∠X in [k] is also preserved synchronously to maintain the structure of the complex spectrum unchanged; the amplitudes of the shielded areas are set to zero while the phase remains unchanged. In this way, a mask view is formed which has significant gaps in the four narrow bandwidths and the rest is consistent with the original spectrum.

[0039] S2.3, Double-view sample assembly: the preprocessed spectrum {X in [k] | k = 0, …, L-1} is regarded as a complete spectrum view, and {X mask [k] | k = 0, …, L-1} is regarded as a mask spectrum view; both are packaged according to the same frame number to form a pair of self-supervised positive sample views <X in , X mask >. In the subsequent contrastive learning stage, all views from the same frame are labeled as "positive pairs", while views from different frames are labeled as "negative pairs" to meet the input format requirements of the contrastive loss function. In this embodiment, the training batch size is fixed at 256 frames, so after collecting 256 pairs of <original spectrum, mask spectrum> continuously, a mini-batch input tensor X

[0040] S3, Self-supervised contrastive learning representation training: input the double-view samples into the feature encoder, train the encoder through the contrastive learning loss function, and freeze the encoder parameters after training to enable the encoder to learn the discriminative features that distinguish signal components from interference components; the process is shown in Figure 3 , and specifically includes:

[0041] S3.1, Feature encoder construction: double-view batch is input into the fixed-structure spectral domain encoder E θ (·). The encoder is composed of L e= 6 layers of one-dimensional depth separable convolution stack, the first layer contains 32 channels, and the number of channels is doubled every level until 1024; the convolution kernel size is constant at three points and both sides are zero padded to maintain the length, the step size is constant at 1; after each convolution, batch normalization and hyperbolic tangent activation function are connected; the parameter set θ uniquely identifies the encoder weight throughout the process, and is no longer reused in other modules.

[0042] Encoder E θ Output fixed dimension embedding vector:

[0043] z = E θ (X in ), z' = E θ (X mask )

[0044] Where z and z' belong to The vector dimension d is the final number of channels of the network; after the convolution stack, the spectral vector is compressed into a real number embedding with length d = 128 And

[0045] S3.2, Definition of contrastive positive samples: In the contrastive learning stage, two views generated in the same frame are used as positive sample pairs. Specifically, if the view index is marked as j (i.e. the same frame), then (z (j) , z ′(j) ) is considered as a "positive pair"; while different frame indexes p ≠ q get embeddings z (p) , z ′(q) are collectively referred to as "negative pairs"; this positive-negative division remains fixed in subsequent loss calculation to avoid cross interference.

[0046] S3.3, Contrastive loss optimization and encoder solidification: For the same batch of data B, first calculate the cosine similarity, then construct the NT-Xent loss:

[0047]

[0048] Where τ is the temperature coefficient, in this embodiment, τ = 0.07; after gradient backpropagation, update θ using stochastic gradient descent with momentum 0.9; the initial learning rate is set to 10 -3 and decay in a cosine annealing manner. When The decline amplitude in the next several rounds of iteration is lower than the preset threshold ε = 10 -4 , it is considered as training convergence; at this time, all encoder parameters θ are frozen, and subsequent steps only perform forward inference without modifying parameters. The freezing action marks the end of the representation learning stage and provides stable and label-independent spectral features for the next step.

[0049] S4, frequency-gated interference suppression: a frequency-gated attention network is introduced to dynamically generate an amplitude-phase dual-channel suppression mask based on the encoder output features, and the mask is multiplied with the preprocessed spectrum to obtain the power spectrum after interference suppression; the frequency-gated attention network includes an amplitude-gated subnetwork and a phase-gated subnetwork, wherein the amplitude-gated subnetwork is used to output an amplitude suppression coefficient, and the phase-gated subnetwork is used to output a phase correction amount; the two together constitute an amplitude-phase dual-channel suppression mask; the process is as shown in Figure 4

[0050] S4.1, amplitude-gated path: input the encoded features into the amplitude-gated subnetwork to output an amplitude suppression coefficient between 0 and 1, which acts on the power spectrum in the frequency dimension;

[0051] the 1024-dimensional feature vector of the output of the frozen encoder at the kth frequency where k is the frequency index; through a one-dimensional depth separable convolution operator

[0052]

[0053] to obtain a channel compression vector; w q is a fixed convolution kernel weight set, and the weights of each channel are independent of each other; ⊙ represents a channel-by-channel product, i.e. Hadamard product;

[0054] then a Sigmoid activation is used:

[0055] g mag [k]=σ(DWConv(h[k]))

[0056] where g mag [k]∈(0,1) is the amplitude suppression coefficient, which describes the proportion of energy to be preserved at the kth frequency; for easy end-side deployment, the convolution weights and the activation results are uniformly quantized to 8-bit unsigned integers in the inference stage.

[0057] S4.2, phase correction path: input the encoded features into the phase-gated subnetwork to output a phase correction amount between -π and π, which is used to compensate for the phase distortion introduced by the interference; a convolution kernel with shared weights with S4.1 is used to process the phase component of the encoder output

[0058] φ[k]=∠X in [k]

[0059] apply a convolution of the same size and pass through a hyperbolic tangent mapping

[0060] g pha [k]=πtanh(DWConv(φ[k]))

[0061] where g pha ​[k] is the phase correction quantity, which is used to offset the phase distortion caused by the trailing interference; the constant π is a fixed scaling factor; the phase compensation quantity can fine-tune the trailing phase introduced by the interference while maintaining the original coherence characteristics, and ensure that the suppressed signal does not appear to be broken. In line with the amplitude branch, all operations in the reasoning stage also use 8-bit fixed-point format.

[0062] S4.3, Gating fusion and spectral filtering: According to the quantities obtained in S4.1 and S4.2, the amplitude suppression coefficient and the phase correction quantity are combined in the frequency domain in an element manner to construct a complex suppression mask:

[0063]

[0064] Where G[k] adjusts the amplitude and phase at the same time; multiply the mask on the preprocessed spectrum:

[0065]

[0066] Get the suppressed spectrum Since g mag [k] suppresses the energy components judged to be interference, and g pha [k] also corrects the corresponding phase, and the suppression process reduces the interference power and maintains the consistency of the amplitude-phase of the useful signal. After completing this multiplication, no further modification is made to the spectrum; the resulting suppression spectrum In the subsequent steps, it is directly sent into the inverse Fourier transform to generate the time-domain signal after interference suppression.

[0067] S5, Time-domain reconstruction and output: the power spectrum after interference suppression is subjected to inverse Fourier transform to obtain the restored time-domain signal; the specific implementation method is: the suppression result in the spectral domain is restored to a continuous time-domain signal; first, the inverse fast Fourier transform is performed on each frame:

[0068]

[0069] Where n is the time-domain sample index, and N is the fixed number of transform points;

[0070] Then the overlap-preservation strategy is used: the adjacent frames maintain a fixed overlap ratio of 0.5 on the time axis, and the first half of the latter frame is added to the second half of the former frame; this ratio runs through all the embodiments and is not replaced. The system continues to use the window configuration of 4096 frame length and 2048 overlap in step 1: the first 2048 points of the transformed half frame are added to the last 2048 points of the residual of the previous frame to realize the overlap-preservation splicing of 0.5, and after frame-level splicing, the continuous sequence m is the global time index. It is directly sent to the demodulation or the physical layer processing without additional filtering. Under the verification of the audio bandwidth monitor, there is no amplitude step or phase misalignment at the splicing boundary; compared with the pre-suppression signal, the link error tester shows that the packet error rate is significantly reduced under the same modulation and demodulation configuration, indicating that the interference power has been effectively weakened and the amplitude-phase integrity of the effective signal is maintained. The processing delay only contains one IFFT and vector addition, and the whole time domain reconstruction-splicing process can be completed in real time inside the single-core embedded DSP, and immediately output to the subsequent synchronous demodulator or beamforming module.

[0071] S6, online adaptation and dynamic update: the application monitors the change of the link signal-to-noise ratio during the operation of the system, adjusts the temperature coefficient and the mask strategy according to the real-time feedback, and updates the attention network with incremental data to continuously optimize the suppression effect. The specific method is: the system evaluates the residual interference power once every fixed time interval ΔT = 1 s during operation:

[0072]

[0073] where X in [k] is the pre-processed spectrum generated in step S1. If P res is higher than the preset threshold P th = -20 dB, trigger light incremental learning: keep the encoder weight frozen, only perform gradient descent on the frequency gating network parameter w once:

[0074]

[0075] where η = 5 × 10-4 is the learning rate, is the contrastive loss defined in step S3. If P res ≤ P th , determine that the suppression effect meets the standard, and the frequency gating network retains the existing weight and is not modified. In this way, the suppression performance can be maintained when the channel conditions change or new interference appears without manual intervention, completing the whole real-time closed loop. The fine-tuning operation can be completed in one forward-backward-update cycle and will not interrupt the data stream; in the experiment, when the environment suddenly appears frequency shift interference or the sweep width suddenly expands, the online update can press the residual power back below the threshold within two to three detection periods, and the parameters also remain static when the interference environment returns to normal, avoiding unnecessary drift.

[0076] In summary, the present application introduces a random band mask-double view self-supervised training framework, combined with amplitude-phase dual channel frequency gating attention mechanism, to construct a spectral domain deep learning method that can suppress wideband unknown interference in real time without manual annotation. The method first learns the discriminant representation of signal and interference under the condition of no label using contrastive loss, and then completes energy suppression and phase compensation in the spectral domain at one time with the gating mask, thereby significantly reducing the interference power in a dynamic wireless environment without destroying the coherence of the useful signal. Compared with traditional methods that rely on pilots, adaptive FIR or supervised classification, the present application completely eliminates the annotation and calibration costs, while avoiding the damage of hardware trap filters to the main lobe of the signal, significantly improving the reliability and resource utilization efficiency of the communication link.

[0077] During training and deployment, the present application uses NT-Xent contrastive loss with fixed temperature, combined with a 6-layer depth separable convolutional encoder, to define the original spectrum and mask spectrum generated from the same frame as a positive sample pair, and to combine across frames as a negative sample pair, thereby strengthening the separability of interference and signal in the feature space. After the encoder converges, it is quantized to 8-bit fixed-point format and frozen, and the gating network retains an online fine-tuning interface; the system monitors the residual interference power every fixed time, and when the suppression effect is below the threshold, it performs a lightweight gradient update on the gating weights, allowing the model to continuously adapt to scenarios such as frequency shift and sweep width changes without human intervention, achieving true edge self-evolution.

[0078] The present application provides a set of efficient, lightweight, and embeddable broadband interference real-time suppression scheme, suitable for satellite communication, cellular base station, Internet of Things terminal, unmanned aerial vehicle link and other application scenarios, especially suitable for complex, variable and labeled data scarce front-line or edge devices. Through software and hardware co-design, this method can achieve millisecond-level inference delay on a single-core DSP or small FPGA, providing a solid technical support for building an intelligent and low operation and maintenance cost wireless transmission system.

[0079] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present application and should not be construed as limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations of the present application without departing from the spirit of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A wideband unknown interference real-time suppression method based on spectral domain self-supervision, characterized in that, The method comprises the following steps: S1, signal acquisition and spectral domain preprocessing: performing fast Fourier transform on the received baseband IQ signal to obtain a power spectrum; and performing baseline correction and amplitude limiting on the spectral data to obtain a preprocessed spectrum; S2, random mask view construction: randomly selecting non-overlapping frequency bands in the preprocessed spectrum domain and performing amplitude masking to generate a mask spectrum; The mask spectrum and the original preprocessed spectrum form a double-view sample; S3, self-supervised contrastive learning representation training: inputting the double-view sample into a feature encoder, training the encoder through a contrastive learning loss function, and freezing the encoder parameters after training; S4, frequency-gated interference suppression: introducing a frequency-gated attention network, generating an amplitude-phase dual-channel suppression mask based on the encoder output feature, and multiplying the mask with the preprocessed spectrum to obtain an interference-suppressed power spectrum; S5, time domain reconstruction and output: performing inverse Fourier transform on the interference-suppressed power spectrum to obtain a restored time domain signal.

2. The method of claim 1, wherein, The S1 specific implementation method is: receiving a baseband IQ signal and performing discrete sampling at a fixed sampling rate Fs, wherein is a time series, and the length of each frame of samples after sampling is L; applying a window function to each frame sample and performing a fast Fourier transform to obtain a complex spectrum ; obtaining a power spectrum then continue to calculate power and use a three-point sliding median estimate the background noise baseline so as to obtain a baseline corrected spectrum ; According to a fixed shear coefficient α∈(0,1), a threshold value is calculated: ； And perform amplitude limiting: ； The final pre-treatment spectrum is formed for use in the next step .

3. The method of claim 2, wherein, The S2 specifically comprises: S2.1, Random band selection: in pre-processing spectrum , first set a positive integer , denoting the number of bands to be shielded; divide the non-negative frequency half-axis into several candidate intervals uniformly; then generate a starting index for each mask band using discrete uniform distribution; set a constant bandwidth P, and construct an index set for each interval; if there is an intersection between some and existing , resample until all are mutually exclusive; finally, regard as the mask index set of the current frame signal; S2.2, Mask spectrum generation: At the index set The binary mask function is defined above: ； The mask spectrum is obtained by using the function: ； and the original phase angle is preserved simultaneously to maintain the complex spectrum structure unchanged; S2.3, Dual-view sample assembly: Assemble pre-processed spectra as complete spectral views, and as masked spectral views; both are packed according to the same frame number, forming a pair of self-supervised positive sample views In the subsequent contrastive learning phase, all views from the same frame are labeled as "positive pairs", while views from different frames are labeled as "negative pairs".

4. The method of claim 1, wherein, The S3 specifically comprises: S3.1, Feature encoder construction: The encoder is constructed by Layer One-dimensional depth separable convolution stack; Encoder Output fixed-dimension embedding vector: ； wherein and belong to , vector dimension is the final number of network channels S3.2, Contrastive positive sample definition: The two views generated in the same frame are taken as the positive sample pair in the contrastive learning stage. If the view index is marked as , then is identified as "positive pair"; while the embedding obtained from different frame indices are collectively referred to as "negative pair"; S3.3, contrastive loss optimization and encoder freezing: for the same batch of data, first calculate the cosine similarity, then construct the NT-Xent loss; after gradient backpropagation, update using stochastic gradient descent ; when the drop in the number of iterations in the next few rounds is lower than the preset threshold , it is considered that the training converges; at this time, all encoder parameters are frozen , and subsequent steps only perform forward inference without modifying the parameters.

5. The method of claim 1, wherein, In the S4, the frequency-gated attention network comprises an amplitude gating sub-network and a phase gating sub-network, the amplitude gating sub-network is used for outputting an amplitude suppression coefficient, and the phase gating sub-network is used for outputting a phase correction amount; and the two together constitute an amplitude-phase dual-channel suppression mask; The process specifically comprises: S4.1, amplitude gating path: inputting the encoded feature into the amplitude gating sub-network to output an amplitude suppression coefficient between 0 and 1, and applying the amplitude suppression coefficient to the power spectrum in the frequency dimension; the frozen encoder outputs a feature vector at the kth frequency by a one-dimensional depth separable convolution operator obtaining a channel compression vector Then, a Sigmoid activation is adopted to obtain the amplitude suppression coefficient; S4.2, phase correction path: inputting the encoded feature into the phase gating sub-network to output a phase correction amount between −π and π; simultaneously, a convolution kernel shared with S4.1 is used to perform a same-size convolution on the phase component output by the encoder and to obtain the phase correction amount through a hyperbolic tangent mapping; S4.3, Gating fusion and spectral filtering: according to the amplitude suppression coefficient and the phase correction quantity obtained in S4.1 and S4.2, the amplitude suppression coefficient and the phase correction quantity are combined in the frequency domain in an element manner to construct a complex suppression mask; the mask is multiplied by the preprocessed spectrum to obtain a suppressed spectrum .

6. The method of claim 1, wherein, The S5 specifically realizes the method as follows: in the spectral domain, the result is inhibited is restored to a continuous time domain signal; first, inverse fast Fourier transform is performed on each frame; Then, an overlap-preservation strategy is adopted: adjacent frames maintain a fixed overlap ratio of 0.5 on the time axis, and the first half of the latter frame is added to the second half of the former frame.

Images