Biomedical signal denoising method and system based on double-branch state space model and phase perception
By using a dual-branch state-space model and the phase-aware SpectroMamba-UNet architecture, the problems of high computational complexity and loss of phase information in EEG signal processing are solved, and efficient and high-fidelity signal reconstruction is achieved on edge devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA ZHONGSHAN INST
- Filing Date
- 2026-03-13
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies for processing EEG signals suffer from high computational complexity, loss of phase information leading to waveform distortion, and difficulty in taking into account time and frequency domain characteristics. Traditional methods are also difficult to perform efficient noise reduction on edge devices.
The SpectroMamba-UNet architecture, based on a dual-branch state-space model and phase perception, is adopted. By processing the temporal branch and the phase-aware spectral branch in parallel and combining the real and imaginary parts splicing strategy, comprehensive modeling and high-fidelity reconstruction of EEG signals can be achieved.
It achieves efficient denoising of EEG signals with linear computational complexity, significantly improving the waveform fidelity and artifact removal capability of signal reconstruction, and is suitable for resource-constrained edge devices.
Smart Images
Figure CN122020003A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of biomedical signal processing and artificial intelligence, and in particular to a method and system for denoising biomedical signals such as electroencephalograms (EEGs) by using a state-space model (Mamba) for bi-branch feature extraction and combining it with a phase-aware strategy. Background Technology
[0002] Electroencephalography (EEG) is an important non-invasive technique for diagnosing neurological disorders such as epilepsy and sleep apnea. However, EEG signals are highly susceptible to physiological artifacts such as electrooculography (EOG) and electromyography (EMG), as well as environmental noise (such as power line interference). These artifacts are usually non-stationary and overlap with neural oscillation signals in the frequency band, making it difficult for traditional filtering methods (such as Wiener filtering and wavelet transform) to maintain signal fidelity while removing noise.
[0003] In recent years, deep learning technology has made significant progress in the field of biomedical signal processing. Convolutional neural networks (CNNs, such as ResNet and U-Net) have become commonly used baseline methods, but they are limited by local convolutional kernels, making it difficult to capture long-range dependencies and global spectral artifacts. While the Transformer architecture solves the global modeling problem through its self-attention mechanism, its... Its quadratic computational complexity makes it inefficient when processing long physiological sequences, making it difficult to deploy on resource-constrained edge devices such as wearable EEG monitors.
[0004] Recently, Structured State-Space Models (SSMs), especially the Mamba architecture, have gained attention due to their linear computational complexity. It has attracted much attention for its ability to effectively model long sequences. However, standard Mamba mainly operates in the time domain and is insufficient in handling artifacts with specific characteristics in the frequency domain. Existing methods that combine frequency domain analysis usually process the amplitude spectrum after the Fast Fourier Transform (FFT) directly, often ignoring phase information. The loss of phase information leads to waveform distortion and time alignment deviations (such as peak drift) in the reconstructed signal, which seriously affects the diagnostic accuracy of clinical features (such as the onset of P waves or spikes).
[0005] Therefore, there is an urgent need for a unified image communication and signal processing method that can maintain linear computational complexity while simultaneously performing accurate modeling in both the time and frequency domains (especially preserving phase information). Summary of the Invention
[0006] The purpose of this invention is to overcome the problems of high computational complexity, waveform distortion caused by loss of phase information, and difficulty in taking into account time and frequency domain characteristics in the prior art, and to provide a biomedical signal denoising method and system (SpectroMamba-UNet) based on a dual-branch state-space model and phase sensing.
[0007] Technical solution
[0008] To achieve the above objectives, the present invention adopts the following technical solution:
[0009] Methodological aspects
[0010] This invention proposes a biomedical signal denoising method based on a dual-branch state-space model and phase sensing. This method constructs a denoising model based on SpectroMamba-UNet. The model employs a U-shaped encoder-decoder architecture, with each layer consisting of a SpectroMamba module. This module includes parallel temporal branches and phase-sensing spectral branches.
[0011] The time-domain branch utilizes a bidirectional Mamba layer to capture the temporal dynamics of the signal;
[0012] The phase-aware spectral branch obtains the complex spectrum through FFT transformation and innovatively employs a "real and imaginary part concatenation strategy," concatenating the real and imaginary parts of the complex spectrum along the channel dimension to generate a real-valued spectral feature tensor. This tensor is then globally modeled using a spectral domain Mamba layer, and finally recovered through separation and IFFT. In this way, the model can simultaneously learn the joint distribution of amplitude and phase, achieving precise suppression of frequency domain artifacts and complete preservation of phase information.
[0013] System aspects
[0014] This invention also provides a biomedical signal denoising system based on a bi-branch state-space model and phase sensing, comprising a signal acquisition module, a model storage and execution module, a SpectroMamba processing unit (including a time-domain processing submodule and a phase-sensing spectrum processing submodule), a feature fusion module, and a signal reconstruction module. The phase-sensing spectrum processing submodule is specifically configured with a real-to-imaginary splicing unit for performing channel splicing operations between the real and imaginary parts.
[0015] Beneficial effects
[0016] Compared with the prior art, the present invention has the following advantages and positive effects:
[0017] This invention introduces a dual-branch SpectroMamba architecture to achieve comprehensive modeling of non-stationary signals. By integrating SpectroMamba modules at each level of a U-shaped network and setting up time-domain and phase-aware spectrum branches in parallel, the model can simultaneously capture the transient changes in the time domain (such as spikes) and global features in the frequency domain (such as artifacts in specific frequency bands) of the signal. This enables deep feature extraction and comprehensive modeling of non-stationary biomedical signals, overcoming the limitations of traditional single-domain processing.
[0018] This invention proposes a phase-aware spectral modeling and real-imaginary concatenation strategy to achieve high-fidelity waveform reconstruction. Addressing the phase loss issue inherent in existing frequency-domain deep learning methods due to direct amplitude spectrum processing, this invention employs a "real-imaginary concatenation strategy." By concatenating the real and imaginary parts of the FFT-transformed signal along the channel dimension and jointly learning them, phase information is explicitly preserved. This design ensures precise alignment of the denoised signal waveform on the time axis, effectively avoiding waveform drift and distortion, and significantly improving the waveform fidelity of the reconstructed signal (Pearson correlation coefficient significantly superior to mainstream models).
[0019] This leverages the linear computational complexity to achieve efficient and lightweight deployment. It utilizes the linear complexity of the state-space model (Mamba). It replaces the quadratic complexity of the traditional Transformer architecture. This computational advantage enables the invention to achieve extremely high computational efficiency and lower memory usage when processing long sequences of biomedical signals. Furthermore, compared to mainstream models of the same type, the invention significantly reduces the number of model parameters while maintaining high performance, thus meeting the real-time edge processing requirements of portable medical devices or wearable EEG monitors.
[0020] Thorough artifact removal and improved robustness in low signal-to-noise ratio (SNR) environments: The dual-branch design and composite loss constraint mechanism of this invention enable the model to simultaneously handle time-domain non-stationary artifacts and frequency-domain specific frequency band artifacts (such as low-frequency drift in electrooculography and high-frequency noise in electromyography). Even in low SNR environments, this invention maintains excellent denoising performance, with its SNR significantly outperforming existing baseline methods. Attached Figure Description
[0021] Figure 1This is a schematic diagram of the overall architecture of the biomedical signal denoising system based on SpectroMamba-UNet according to the present invention. As shown in the figure, the data flow sequentially passes through the input signal interface (100), the encoder layer (101-104) composed of SpectroMamba modules, the bottleneck layer module (105), and the decoder layer (106-109) along the direction of the arrows, and finally reaches the output signal interface (110). The long-span dashed arrows in the figure represent skip connections that fuse the features of the encoder layer to the corresponding decoder layer.
[0022] Figure 2 This is a schematic diagram of the internal processing flow of the SpectroMamba module, the core component of this invention. As shown in the figure, the flow details the processing steps of the input features (200) after they are distributed through the branch path in the time domain branch (201-202) and the frequency domain branch (203-206), as well as the final feature fusion (207) and post-processing (208-209) processes.
[0023] Figure 3 This is a schematic diagram of the data processing flow for the real and imaginary part splicing strategy in the phase-aware spectral branch of this invention. As shown in the figure, the complex spectrum (300) is decomposed into a real part (301) and an imaginary part (302) by the distribution node, and then a real-valued spectral feature tensor is generated through splicing operation (303) for spectral domain Mamba modeling (304). The modeled data is separated into the recovered real part (305) and imaginary part (306) by the reconstruction node, and finally reassembled and inversely transformed into time-domain features (307).
[0024] Figure 4 This image shows a comparison of the denoising performance of the SpectroMamba-UNet-based denoising method of this invention in simulation experiments. Sub-image (a) is a comparison of time-domain waveform reconstruction, showing the change in signal amplitude over time; sub-image (b) is a comparison of power spectral density (PSD), showing the distribution of signal energy in the frequency domain. Different line types (dark gray solid line, black wide dashed line, black dotted line) distinguish the noisy input, the original true value, and the output of this invention. Detailed Implementation
[0025] The preferred embodiments of the present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited thereto.
[0026] Example 1: Noise Reduction Method
[0027] like Figure 1 As shown, the denoising method based on SpectroMamba-UNet of the present invention includes the following steps:
[0028] Step 1: Data Acquisition and Input The noisy biomedical signal sequence to be processed is acquired and input into the system via the input signal interface (100). This signal is typically modeled as a superposition of clean signal and additive noise. To stabilize the training dynamics, the input signal is normalized.
[0029] Step 2: Model Construction and Forward Propagation. Construct the SpectroMamba-UNet denoising model. For example... Figure 1 As shown, the model logically follows a U-shaped encoding-decoding structure, and its physical processing flow is as follows:
[0030] Encoding stage (downsampling path): The signal passes through four cascaded SpectroMamba modules (101, 102, 103, 104) sequentially. Each module extracts deep semantic features and gradually reduces the temporal resolution of the features.
[0031] Bottleneck stage: The encoded features enter the SpectroMamba module (105) in the middle layer. Here, the feature dimension is the highest and the receptive field is the largest, which is used to capture global context information.
[0032] Decoding stage (upsampling path): The data then enters four cascaded SpectroMamba modules (106, 107, 108, 109) for feature reconstruction.
[0033] Feature fusion (skip connections): such as Figure 1 As shown by the dashed arrows, the model introduces a skip connection mechanism. Specifically, the output features of module (101) are fused to module (109) via dashed connections; module (102) is fused to module (108); module (103) is fused to module (107); and module (104) is fused to module (106). This design directly passes the high-resolution detail information from shallow layers to the decoder, solving the problem of detail loss in deep networks.
[0034] Output stage: The features processed by module (109) are finally reconstructed into clean biomedical signals through the output signal interface (110).
[0035] Step 3: SpectroMamba module processing
[0036] like Figure 2 As shown, the detailed processing flow of the SpectroMamba module for input features is as follows:
[0037] Feature Input and Distribution: The features output from the previous level are used as module input features (200) and enter the current module. Subsequently, the data stream enters the path distribution node and is transmitted in parallel to the time domain branch on the left and the frequency domain branch on the right.
[0038] Temporal branch processing (left path): The input features first pass through the LayerNorm module (201) to stabilize the feature distribution and accelerate model convergence.
[0039] The normalized features are fed into the bidirectional Mamba layer (202). This module contains two independent State-Space Model (SSM) submodules, which perform forward and backward scans respectively. The forward scan submodule updates the hidden states sequentially at time steps t=1,2,...,N, using the following recursive formula: Where A and B are learnable state transition matrices, Enter the current time. This is the current hidden state; the output of this submodule. The entire historical context information prior to each time step is captured. The backward scan submodule flips the input sequence along the time axis and recursively processes it in the same way, effectively capturing the future context information after each time step. The outputs of the two submodules are concatenated along the channel dimension and linearly projected to obtain the final output, thus achieving complete modeling of the bidirectional context dependency of long sequence signals. Since the computational cost of each SSM submodule is linearly related to the sequence length N, i.e., each time step only performs fixed-dimensional matrix-vector multiplication without interacting with the entire sequence, the total computational complexity of the bidirectional structure remains O(N), significantly better than the Transformer self-attention mechanism. This achieves quadratic complexity, thus capturing the bidirectional contextual dependencies of long sequence signals with linear computational complexity.
[0040] Frequency domain branch processing (right path): The input features are first transformed from the time domain to the frequency domain by the Fast Fourier Transform (FFT) module (203) to generate a complex spectrum.
[0041] The process then proceeds to the real-imaginary part splicing module (204). In this step, the real and imaginary parts of the complex spectrum are separated and spliced along the channel dimension to generate a real-valued spectral feature tensor with double the number of channels.
[0042] spliced real-valued spectral feature tensor The input spectral domain Mamba module (205) is used for global frequency dependency modeling and phase-amplitude joint learning. Specifically, the spectral domain Mamba module will... Modeling the sequence sequentially along the frequency index dimension as a one-dimensional sequence: The k-th position simultaneously contains the real part of the k-th frequency component. and imaginary part value ,Right now The amplitude of this frequency component (i.e., the signal energy contribution) is The phase angle is The state equations of the spectral domain Mamba modules are passed sequentially between frequency positions through hidden states:
[0043] ,
[0044] This approach allows the amplitude information (energy) at each frequency position to be processed jointly with its corresponding frequency index position (phase reference coordinate) during the same state update process, thus explicitly establishing the correlation between signal energy and phase position. Since both real and imaginary parts are used as input, the model can learn the nonlinear coupling characteristics of amplitude and phase in real-valued space, avoiding the problem of phase information loss when traditional methods only process the amplitude spectrum.
[0045] After the modeled features are separated and recombined, they are restored to time-domain features using the Inverse Fast Fourier Transform (IFFT) module (206). The specific steps are as follows: The real-valued spectral feature tensor output by the spectral domain Mamba module... It is equally divided into a first half and a second half along the channel dimension, which are respectively used as the real parts of the reconstructed complex spectrum. and the virtual part Recombining the two into a complex spectrum ; then on Perform an inverse fast Fourier transform:
[0046]
[0047] Mapping the frequency domain complex representation back to the time domain and outputting the real-valued time domain feature sequence completes the full processing loop of the phase-aware spectrum branch from the frequency domain to the time domain.
[0048] Feature fusion: Temporal branch output (From 202) and frequency domain branch output (From 206) Import the element-wise addition module (207) to achieve the superposition and fusion of dual-domain features:
[0049]
[0050] Post-processing and output: The fused features pass through a random dropout and residual connection module (208). Random dropout is used to prevent overfitting, while residual connection passes the input features to the module. (200) is added to the current result to alleviate the vanishing gradient problem. The final output is:
[0051]
[0052] As feature output (209) is passed to the next level of the network.
[0053] like Figure 3 As shown, in order to explicitly preserve phase information in frequency domain processing, the present invention designs the following specific data transfer steps:
[0054] Spectrum generation: After the input feature x is transformed by FFT, a complex spectrum Z(300) is generated:
[0055]
[0056] This complex signal also contains the original amplitude information. With phase information .
[0057] Separation and splicing of real and virtual elements:
[0058] The system first performs parallel decomposition processing on the complex number Z, performing real part extraction (301) and imaginary part extraction (302) respectively.
[0059] The extracted real and imaginary features are then fed into the real-imaginary concatenation module (303), where they are concatenated along the channel dimension to generate a real-valued feature tensor.
[0060]
[0061] This transforms complex field information into a high-dimensional real-valued feature tensor.
[0062] spliced real-valued spectral feature tensor It is fed into the spectral domain Mamba module (304). This module models the global dependencies of frequency components in the real-valued space, and its state recursion is as follows:
[0063] ,
[0064] Due to input Including both real and imaginary part information, the model can jointly learn the nonlinear coupling characteristics of amplitude and phase.
[0065] Output tensor after modeling Separate reconstruction is performed to obtain the recovered real parts. and restore the virtual part Finally, the restored real and imaginary parts are recombine into a complex form (307). The complex spectrum of this recombination Then perform the IFFT operation:
[0066]
[0067] Restored to time-domain feature sequence This completes the closed-loop processing of the frequency domain branch.
[0068] Example 2: Noise Reduction Effect Analysis
[0069] To verify the effectiveness of the proposed SpectroMamba-UNet model, a simulated signal denoising experiment was conducted. The experimental results are as follows: Figure 4 As shown in the figure, different combinations of gray levels and line styles are used to distinguish different signals. The specific analysis is as follows:
[0070] Time-domain waveform reconstruction analysis (e.g.) Figure 4 (a) shown)
[0071] The dark gray solid line in the figure represents the noisy input signal. It can be seen that it is contaminated by a large amount of low-frequency artifacts (simulating EOG drift), and the waveform fluctuates violently, which masks the real EEG characteristics.
[0072] The wide black dashed line in the figure represents the original clean signal without noise (Ground Truth), which is the target baseline we expect to recover.
[0073] The black dotted lines (long dash-dot-long dash) in the figure represent the noise-reduced output signal (Output) processed by the system of this invention.
[0074] Comparison results: From Figure 4 As can be clearly observed in (a), the black dotted line (output of this invention) and the wide black dashed line (original true value) exhibit a very high degree of overlap across the entire time axis (1.5s-3.5s). This indicates that the present invention successfully removes the large artifacts shown by the dark gray solid line, while accurately preserving the minute fluctuation details of the signal, without significant amplitude attenuation or phase shift.
[0075] Frequency domain power spectral density analysis (e.g.) Figure 4 (b)
[0076] Figure 4 (b) shows the power spectral density (PSD) distribution of the corresponding signal, with the horizontal axis representing frequency and the vertical axis representing the power spectral density on a logarithmic scale.
[0077] Low-frequency region (0-5Hz): The dark gray solid line in this region has extremely high energy, corresponding to artifact drift in the time domain. After processing by this invention, the energy of the black dotted line in this frequency band is significantly reduced and closely follows the wide black dashed line (the true value), demonstrating the model's excellent ability to suppress low-frequency artifacts.
[0078] Characteristic frequency bands (10Hz and 20Hz): At the characteristic peaks of 10Hz (Alpha wave) and 20Hz (Beta wave), the peak height of the black dotted line is exactly the same as that of the wide black dashed line. This indicates that the model perfectly preserves the key spectral characteristics of biomedical signals while removing noise, without losing useful information.
[0079] In summary, such as Figure 4 As shown by the height fit between the midpoint line and the wide dashed line, this invention can effectively separate complex background noise and artifacts in biomedical signals, and achieve high-fidelity signal reconstruction in both the time and frequency domains.
[0080] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. A biomedical signal denoising method based on a bi-branch state-space model and phase sensing, characterized in that, include: A noisy biomedical signal sequence to be processed is acquired; a denoising model based on SpectroMamba-UNet is constructed, wherein the model adopts a U-shaped encoder-decoder architecture, and each level of the encoder and decoder is composed of SpectroMamba modules; the noisy biomedical signal sequence is input into the denoising model, and feature extraction and reconstruction are performed through the SpectroMamba modules, wherein the SpectroMamba modules include parallel temporal branches and phase-aware spectral branches; The time-domain branch uses a bidirectional Mamba layer to capture the temporal dynamic features of the signal; the phase-aware spectrum branch performs the following steps: performing a Fast Fourier Transform on the input features to obtain a complex spectrum; using a real-imaginary part concatenation strategy, concatenating the real and imaginary parts of the complex spectrum along the channel dimension to generate a real-valued spectral feature tensor; using a spectral domain Mamba layer to globally model the real-valued spectral feature tensor to learn the joint distribution of amplitude and phase; finally, separating the modeled features and restoring them to time-domain features through an inverse Fast Fourier Transform; fusing the temporal dynamic features output by the time-domain branch with the time-domain features output by the phase-aware spectrum branch to obtain a denoised feature representation; and the denoising model outputs a denoised clean biomedical signal based on the feature representation.
2. The method according to claim 1, characterized in that: The specific steps for splicing and processing the real and imaginary parts in the phase-sensing spectral branch include: assuming the input feature is... Through formula Obtaining the complex spectrum Extract the complex spectrum respectively. real part and the virtual part Construct the input tensor Where Concat represents the channel concatenation operation, making The number of channels is the input feature Twice the number of channels; directly applying the real-valued tensor to the spectral domain Mamba layer. The signal energy and phase position are correlated through processing. The processed output tensor is separated in the channel dimension and used as the real and imaginary parts of the reconstructed complex spectrum.
3. The method according to claim 1, characterized in that: The time-domain branch includes a layer normalization layer and a bidirectional Mamba layer connected in sequence; the bidirectional Mamba layer contains a forward scan path and a backward scan path, which are used to capture the bidirectional context dependency of long sequence signals with linear computational complexity.
4. The method according to claim 1, characterized in that: The feature fusion steps of the SpectroMamba module specifically include: adding the temporal dynamic features output by the time-domain branch to the time-domain features output by the phase-aware spectrum branch element by element; performing random deactivation processing on the result after addition; introducing residual connections, and adding the result after random deactivation processing to the input features of the SpectroMamba module again to obtain the final output features.
5. The method according to claim 1, characterized in that: The denoising model also includes skip connections, which are used to fuse the features output by the encoder level to the corresponding decoder level through long-span connections, so as to preserve the high-resolution detail information of the signal.
6. The method according to claim 1, characterized in that: The training process of the denoising model uses a composite loss function. The optimized formula is as follows: in, To predict the L1 loss of the signal compared to the real signal in the time domain, To predict the L1 loss between the signal spectral amplitude and the actual signal spectral amplitude, This is the balance coefficient.
7. The method according to claim 1, characterized in that: The noisy biomedical signal sequence includes electroencephalogram (EEG) signals, and the denoising model is used to remove electrooculogram (EOG) or electromyogram (EMG) artifacts mixed in the EEG signals.
8. A biomedical signal denoising system based on a bi-branch state-space model and phase sensing, characterized in that, include: The signal acquisition module is used to acquire the noisy biomedical signal sequence to be processed; The model storage and execution module stores and runs the SpectroMamba-UNet denoising model, which has a U-shaped architecture and each layer contains SpectroMamba processing units. The SpectroMamba processing units are integrated into the denoising model and are used to perform bi-branch feature extraction. Specifically, they include: a time-domain processing submodule configured with bidirectional Mamba layers to capture the temporal dynamic features of the signal; and a phase-aware spectrum processing submodule for performing frequency-domain modeling, which further includes an FFT transform unit, a real-imaginary concatenation unit, a spectral domain Mamba unit, and an IFFT restoration unit. The FFT transform unit converts the input features into a complex spectrum; the real-imaginary concatenation unit separates the real and imaginary parts of the complex spectrum and performs channel concatenation to generate real-valued spectral features; the spectral domain Mamba unit performs global modeling of the real-valued spectral features to preserve phase information; and the IFFT restoration unit transforms the modeled features back to the time domain. A feature fusion module fuses the outputs of the time-domain processing submodule and the phase-aware spectrum processing submodule. The signal reconstruction module is used to output a denoised clean biomedical signal based on the fused features.
9. The system according to claim 8, characterized in that: The real-to-virtual splicing unit is configured to enable the spectral domain Mamba unit to simultaneously learn the amplitude distribution and phase alignment information of the spectrum by converting the complex spectrum into a real-valued tensor with double channels.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method of any one of claims 1 to 7.