Search

CN-122020003-A - Biomedical signal denoising method and system based on double-branch state space model and phase perception

CN122020003ACN 122020003 ACN122020003 ACN 122020003ACN-122020003-A

Abstract

The invention relates to the technical field of biomedical signal processing and artificial intelligence, and discloses a biomedical signal denoising method and system based on a double-branch state space model and phase perception. The method constructs a SpectroMamba-UNet-based denoising model, adopts a U-shaped architecture, integrates SpectroMamba modules at each level, and models non-stationary signals through parallel time domain branches and phase perception spectrum branches. The phase sensing spectrum branch adopts a real-part and imaginary-part splicing strategy, the real part and the imaginary part of the complex spectrum are spliced in the channel dimension and are jointly learned, the phase information is reserved explicitly, and the waveform drift problem of the traditional frequency domain method is solved. The invention replaces the secondary complexity of the traditional transducer with the linear complexity of the state space model, realizes the accurate removal of the artifacts such as the electrooculogram, the myoelectricity and the like and the high-fidelity reconstruction of the signal waveform, and is suitable for the real-time signal processing of the portable medical equipment.

Inventors

  • LI SHUQIANG
  • GUO JIANPING
  • WANG BAOFU
  • CHEN WEI

Assignees

  • 电子科技大学中山学院

Dates

Publication Date
20260512
Application Date
20260313

Claims (10)

  1. 1. A biomedical signal denoising method based on a dual-branch state space model and phase perception is characterized by comprising the steps of obtaining a to-be-processed noisy biomedical signal sequence, constructing a SpectroMamba-UNet-based denoising model, constructing a U-shaped encoder-decoder framework by the model, inputting the noisy biomedical signal sequence into the denoising model through a SpectroMamba module, carrying out feature extraction and reconstruction through the SpectroMamba module, wherein the SpectroMamba module comprises parallel time domain branches and phase perception spectrum branches, capturing time dynamic features of signals through the time domain branches by utilizing a two-way Mamba layer, carrying out fast Fourier transform on the input features to obtain a complex spectrum, splicing a real part and an imaginary part of the complex spectrum in a channel dimension to generate a real-value spectrum feature tensor, carrying out global modeling on the real-value spectrum feature tensor through a spectral domain Mamba layer to learn the joint distribution of amplitude and phase, finally separating the modeled features, restoring the features into time domain fast Fourier transform to obtain time dynamic features of the signals, carrying out inverse transform on the time domain branches, and carrying out the inverse transform on the time domain features of the obtained by the phase perception spectrum branches, and carrying out the inverse transform on the obtained time domain features.
  2. 2. The method of claim 1, wherein the real-imaginary part splicing and processing steps in the phase-aware spectral branches include providing input features as By the formula Obtaining complex frequency spectrum Respectively extracting the complex frequency spectrums The real part of (2) And imaginary part Constructing an input tensor Wherein Concat denotes a channel splicing operation such that The number of channels is the input feature Twice the number of channels, directly mapping real value tensor by using the layer Mamba of the spectrum domain And separating the processed output tensor in the channel dimension to be respectively used as a real part and an imaginary part of the reconstructed complex frequency spectrum.
  3. 3. The method of claim 1, wherein the time domain branches comprise a layer normalization layer and a bi-directional Mamba layer connected in sequence, and wherein the bi-directional Mamba layer comprises a forward scan path and a backward scan path for capturing bi-directional context dependencies of long sequence signals at linear computational complexity.
  4. 4. The method of claim 1, wherein the feature fusion step of the SpectroMamba module specifically includes adding the time dynamic feature of the time domain branch output and the time domain feature of the phase-aware spectrum branch output element by element, performing random inactivation on the added result, introducing residual connection, and adding the random inactivation result and the input feature of the SpectroMamba module again to obtain a final output feature.
  5. 5. The method of claim 1, wherein the denoising model further comprises a skip connection for fusing features of the encoder level output to the corresponding decoder level over a long span connection to preserve high resolution detail information of the signal.
  6. 6. The method of claim 1, wherein the training of the denoising model uses a composite loss function Optimizing, wherein the formula is as follows: Wherein, the To predict the L1 loss of the signal and the real signal in the time domain, To predict the L1 loss between the signal spectral amplitude and the true signal spectral amplitude, Is a balance coefficient.
  7. 7. The method of claim 1, wherein the noisy biomedical signal sequence comprises an electroencephalogram signal and the denoising model is used for removing mixed electro-oculogram artifacts or myoelectric artifacts in the electroencephalogram signal.
  8. 8. A biomedical signal denoising system based on a double-branch state space model and phase perception is characterized by comprising a signal acquisition module, a model storage and execution module, a SpectroMamba processing unit, a phase perception spectrum processing submodule, an FFT conversion unit, a real and virtual splicing unit, a spectral domain Mamba unit and an IFFT restoration unit, wherein the signal acquisition module is used for acquiring a noise biomedical signal sequence to be processed, the model storage and execution module is used for storing and running SpectroMamba-UNet denoising models which are of U-shaped structures and contain SpectroMamba processing units in each level, the SpectroMamba processing unit is integrated in the denoising models and is used for executing double-branch feature extraction, the time domain processing submodule is configured with a bidirectional Mamba layer and is used for capturing time dynamic features of signals, the phase perception spectrum processing submodule is used for executing frequency domain modeling, the FFT conversion unit is used for converting input features into complex frequency spectrums, the real and virtual splicing unit is used for separating real parts and imaginary parts of the complex frequency spectrums and conducting channel splicing to generate real-value spectrum features, the spectral domain Mamba is used for conducting global modeling on the real-value spectral features to preserve phase information, the time domain restoration unit is used for conducting time domain reconstruction feature extraction, the time domain restoration unit is used for conducting frequency domain modeling, the frequency domain reconstruction module is used for conducting frequency domain modeling, the frequency domain modeling module is used for conducting frequency domain modeling, the frequency domain transformation module is used for carrying inverse domain transformation, and the inverse transform module is used for carrying out inverse transform module, and the phase transform module is used for carrying out phase transform module.
  9. 9. The system of claim 8, wherein the real-imaginary stitching unit is configured to enable the spectral domain Mamba unit to learn the amplitude distribution and phase alignment information of the spectrum simultaneously by converting the complex spectrum into real-valued tensors of double channels.
  10. 10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method of any of claims 1 to 7.

Description

Biomedical signal denoising method and system based on double-branch state space model and phase perception Technical Field The invention relates to the technical field of biomedical signal processing and artificial intelligence, in particular to a method and a system for denoising biomedical signals such as electroencephalogram (EEG) by utilizing a state space model (Mamba) to extract double-branch characteristics and combining a phase sensing strategy. Background Electroencephalogram (EEG) is an important non-invasive technique for diagnosing neurological diseases such as epilepsy and sleep apnea. However, EEG signals are extremely vulnerable to physiological artifacts such as electro-oculography (EOG), myoelectricity (EMG), and the like, and contamination by environmental noise (e.g., power frequency interference). These artifacts are typically non-stationary and overlap the neural oscillation signal in frequency bands, making it difficult for conventional filtering methods (e.g., wiener filtering, wavelet transformation) to remove noise while maintaining signal fidelity. In recent years, deep learning techniques have made remarkable progress in the field of biomedical signal processing. Convolutional neural networks (CNNs, such as ResNet and U-Net) have become a common baseline approach, but are limited by local convolution kernels, which are difficult to capture long-range dependencies and global spectral artifacts. Although the transducer architecture solves the global modeling problem through a self-attention mechanism, itThe secondary computational complexity of (c) makes it inefficient in processing long physiological sequences, difficult to deploy on resource-constrained edge devices such as wearable EEG monitors. Recently, structured State Space Models (SSM), particularly Mamba architecture, have been proposed due to their linear computational complexityAnd is of great interest in efficiently modeling long sequences. However, the standard Mamba operates primarily in the time domain, and lacks the artifact handling capability for having specific features in the frequency domain. The existing method combined with frequency domain analysis generally directly processes the amplitude spectrum after Fast Fourier Transform (FFT), and often ignores the phase information. Loss of phase information can result in waveform distortion and time pairs Ji Piancha (e.g., peak drift) of the reconstructed signal, which severely affects the diagnostic accuracy of the clinical features (e.g., P-wave or spike onset). Therefore, there is a need for a unified image communication and signal processing method that can maintain linear computational complexity and accurately model in both the time and frequency domains (particularly, preserving phase information). Disclosure of Invention The invention aims to solve the problems of waveform distortion, difficulty in considering time-frequency domain characteristics and the like caused by high calculation complexity and phase information loss in the prior art, and provides a biomedical signal denoising method and system (SpectroMamba-UNet) based on a double-branch state space model and phase perception. Technical proposal In order to achieve the above purpose, the invention adopts the following technical scheme: Method aspect The invention provides a biomedical signal denoising method based on a double-branch state space model and phase perception, which constructs a SpectroMamba-UNet-based denoising model. The model adopts a U-shaped encoder-decoder architecture, and the core of the model is that each level consists of SpectroMamba modules. The module comprises parallel time domain branches and phase-aware spectrum branches: The time domain branch captures the time dynamic characteristics of the signal by using a bidirectional Mamba layer; The phase sensing spectrum branches acquire complex frequency spectrums through FFT conversion, a real and imaginary part splicing strategy is innovatively adopted, real and imaginary parts of the complex frequency spectrums are spliced in channel dimensions to generate real-value spectrum characteristic tensors, global modeling is carried out by utilizing a spectrum domain Mamba layer, and finally separation and IFFT recovery are carried out. By means of the method, the model can learn the joint distribution of the amplitude and the phase at the same time, and accurate suppression of frequency domain artifacts and complete reservation of phase information are achieved. System aspects The invention also provides a biomedical signal denoising system based on the double-branch state space model and the phase perception, which comprises a signal acquisition module, a model storage and execution module, spectroMamba processing units (comprising a time domain processing sub-module and a phase perception spectrum processing sub-module), a feature fusion module and a signal reconstruction module. The phase perception spectrum processing submodule