Differential filter and wavelet analysis based method and apparatus for processing neural and physiological signals

Abstract

The method and device for processing neural and physiological signals by differentiable filtering and wavelet analysis can automatically adjust and optimize key parameters such as filter center frequency and bandwidth, significantly improve noise robustness, dynamic adaptability and end-to-end processing efficiency, and are particularly suitable for high-demand scenarios such as brain-computer interface, auxiliary diagnosis of neurological diseases and real-time physiological signal monitoring. The method comprises: (1) window segmentation of the input signal, suppression of power frequency and specific frequency band interference by a differentiable band-stop filtering module, and adaptive extraction of the target frequency band by a band-pass filtering layer with a learnable kernel function to achieve noise suppression and signal enhancement; (2) inputting the preprocessed signal into a time-frequency analysis module, using a parameterized wavelet transform layer to generate an optimal wavelet kernel in real time, combining complex wavelet to retain phase information, and realizing adaptive optimization of time-frequency resolution and feature representation; (3) fusing features through a multi-scale parallel convolution layer, and outputting results by a lightweight classification network.

Description

Technical Field

[0001] This invention relates to the field of biomedical signal processing technology, and more particularly to a method for neural and physiological signal processing using differentiable filtering and wavelet analysis, as well as a device for neural and physiological signal processing using differentiable filtering and wavelet analysis. Background Technology

[0002] Electroencephalography (EEG), electrocardiography (ECG), electromyography (EMG), and magnetoencephalography (MEG) are important data sources for clinical diagnosis, neuroscience research, and brain-computer interfaces, offering advantages such as non-invasiveness and high temporal resolution. However, these signals generally suffer from weak amplitude, strong non-stationarity, and susceptibility to various noises and artifacts. For example, environmental power frequency interference, EEG and EMG artifacts, and equipment noise can significantly reduce the signal-to-noise ratio, posing serious challenges to the accuracy and stability of feature extraction and subsequent analysis.

[0003] Neural signals, represented by EEG, exhibit strong individual variability and dynamic changes in their frequency band distribution and interference types, placing higher demands on the adaptability of filter parameters and the accuracy of transient feature extraction. Traditional signal processing workflows typically employ a multi-step approach: first, using a band-stop filter with fixed parameters (such as a notch filter) to suppress power frequency interference; then, combining independent component analysis or principal component analysis to separate and remove artifact signals; next, using techniques such as bandpass filtering, wavelet transform, or short-time Fourier transform to extract specific frequency bands or time-frequency features; and finally, using classifiers such as support vector machines to complete the analysis and discrimination. Although the above methods have improved the accuracy of signal analysis to some extent, they also have the following limitations: (1) The design of filter parameters and feature extraction operators such as wavelet basis functions is highly dependent on empirical presets, making it difficult to dynamically adapt to signal changes in different subjects and complex application scenarios; (2) Multi-step separation processing may introduce cumulative errors and feature redundancy in each stage of the signal, increasing the computational complexity and end-to-end delay of the overall system, making it difficult to meet the requirements of low latency and high robustness for real-time brain-computer interfaces and clinical monitoring; (3) Heterogeneous toolchains (such as the separation of traditional MATLAB preprocessing and deep learning frameworks) cause data flow interruptions and additional input / output overhead, which is not conducive to system integration and automation.

[0004] In recent years, deep learning technology has brought new opportunities to physiological and neural signal processing, enabling the improvement of system adaptability and classification performance through end-to-end feature learning. However, existing deep learning models either directly extract features from the raw, unprocessed time-domain signals, lacking explicit suppression of noise such as power frequency interference and artifacts; or they rely on traditional preprocessing before inputting into the deep network, still suffering from problems such as parameter rigidity and toolchain fragmentation. Therefore, existing technologies still need improvement in terms of filtering adaptability, time-frequency feature co-optimization, and processing efficiency.

[0005] Currently, the processing of physiological signals such as electroencephalograms (EEGs) mostly adopts a step-by-step process, including fixed-parameter filters to remove noise and artifacts, manual feature extraction using time-frequency analysis tools, and shallow model classification such as support vector machines. These methods rely on expert experience and are difficult to adapt to individual differences and dynamic noise. Manual feature extraction is also subjective and one-sided, and multi-step processing is prone to introducing errors and increasing latency.

[0006] In recent years, deep learning has been introduced into EEG signal processing. End-to-end architectures such as convolutional neural networks and recurrent neural networks have improved feature extraction efficiency, but they lack adaptive mechanisms for noise and dynamic frequency bands.

[0007] The relevant existing technologies mainly include: (1) The wavelet filtering method with dynamic parameter optimization (such as CN 202411982003 A) automatically removes electrooculogram artifacts by training the wavelet decomposition layer and threshold strategy, combined with blind source separation technology, and achieves partial automated noise reduction. However, the filtering parameters and wavelet basis selection still depend on prior settings and lack end-to-end collaborative optimization.

[0008] (2) Based on the adaptive time-frequency feature collaborative framework (such as CN 202510113366 A), the dynamic residual network is used to extract time-domain features and frequency-domain information is weighted and fused through the attention mechanism to enhance the feature representation ability of non-stationary signals. However, the frequency-domain feature extraction and noise suppression process is still decoupled from the deep model and cannot dynamically respond to signal changes.

[0009] (3) End-to-end trainable filter-classification models (such as CN 202411802561 A) dynamically adjust feature weights through recurrent neural networks to optimize long-term dependency modeling, but lack adaptive control for specific frequency bands.

[0010] In summary, existing technologies still have limitations in adaptive filtering, feature learning collaboration, and real-time processing. Summary of the Invention

[0011] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a neural and physiological signal processing method based on differentiable filtering and wavelet analysis, which can realize the automatic adjustment and optimization of key parameters such as filter center frequency and bandwidth, significantly improve noise robustness, dynamic adaptability and end-to-end processing efficiency, and is particularly suitable for high-requirement scenarios such as brain-computer interface, auxiliary diagnosis of neurological diseases and real-time physiological signal monitoring.

[0012] The technical solution of this invention is: a method for neural and physiological signal processing based on differentiable filtering and wavelet analysis, comprising the following steps: (1) The input signal is windowed, and the power frequency and specific frequency band interference are suppressed by the differentiable band-stop filter module. The target frequency band is adaptively extracted by the bandpass filter layer with learnable kernel function to achieve noise suppression and signal enhancement. (2) Input the preprocessed signal into the time-frequency analysis module, use the parameterized wavelet transform layer to generate the optimal wavelet kernel in real time, and combine the complex wavelet to retain the phase information to achieve adaptive optimization of time-frequency resolution and feature representation; (3) Features are fused through multi-scale parallel convolutional layers and the results are output using a lightweight classification network.

[0013] This invention employs a dynamic filtering design using symmetric kernel functions to directly implement notch filtering and bandpass filtering within convolutional layers via cross-correlation operations. This avoids redundant computations involving convolutional kernel flipping in traditional filtering algorithms, significantly reducing computational complexity. A time-frequency analysis module based on a trainable symmetric wavelet kernel adaptively optimizes the bandwidth factor and center frequency, enabling dynamic adjustment of time-frequency resolution and accurately capturing transient features in neural and physiological signals. Leveraging native operators within deep learning frameworks, this invention achieves end-to-end integration of filtering, time-frequency transformation, and classification, supporting efficient inference on accelerated platforms such as GPUs. This significantly reduces latency and eliminates the overhead of cross-platform toolchain switching in traditional step-by-step processes. This solution significantly improves real-time performance and deployment convenience while ensuring signal processing interpretability, meeting the demands of high-real-time scenarios such as brain-computer interfaces.

[0014] A neural and physiological signal processing device based on differentiable filtering and wavelet analysis is also provided, comprising: The signal segmentation module divides the acquired neural or physiological signals into segments of equal length according to a fixed time window. Differentiable band-stop filter module integrates a trainable symmetric impulse response filter kernel in a one-dimensional convolutional layer for adaptive suppression of interference in a specified frequency band and its harmonic components in a signal; The bandpass filtering module adaptively extracts target frequency band components relevant to the analysis task from the raw neural or physiological signals, and introduces a trainable symmetric impulse response filtering kernel into the one-dimensional convolutional layer; The time-frequency analysis module extracts time-frequency domain features from preprocessed neural or physiological signals, enabling precise characterization of transient changes and local frequency components of the signal. Multi-scale parallel convolutional layers and classification networks input the output feature sequences into multiple convolutional channels, each channel using a different convolutional kernel size or receptive field to extract features from the signal; each convolutional channel works in parallel to capture multi-scale information of the signal. Attached Figure Description

[0015] Figure 1A flowchart of the neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to the present invention is shown. Detailed Implementation

[0016] like Figure 1 As shown, a neural and physiological signal processing method based on differentiable filtering and wavelet analysis includes the following steps: (1) The input signal is windowed, and the power frequency and specific frequency band interference are suppressed by the differentiable band-stop filter module. The target frequency band is adaptively extracted by the bandpass filter layer with learnable kernel function to achieve noise suppression and signal enhancement. (2) Input the preprocessed signal into the time-frequency analysis module, use the parameterized wavelet transform layer to generate the optimal wavelet kernel in real time, and combine the complex wavelet to retain the phase information to achieve adaptive optimization of time-frequency resolution and feature representation; (3) Features are fused through multi-scale parallel convolutional layers and the results are output using a lightweight classification network.

[0017] This invention employs a dynamic filtering design using symmetric kernel functions to directly implement notch filtering and bandpass filtering within convolutional layers via cross-correlation operations. This avoids redundant computations involving convolutional kernel flipping in traditional filtering algorithms, significantly reducing computational complexity. A time-frequency analysis module based on a trainable symmetric wavelet kernel adaptively optimizes the bandwidth factor and center frequency, enabling dynamic adjustment of time-frequency resolution and accurately capturing transient features in neural and physiological signals. Leveraging native operators within deep learning frameworks, this invention achieves end-to-end integration of filtering, time-frequency transformation, and classification, supporting efficient inference on accelerated platforms such as GPUs. This significantly reduces latency and eliminates the overhead of cross-platform toolchain switching in traditional step-by-step processes. This solution significantly improves real-time performance and deployment convenience while ensuring signal processing interpretability, meeting the demands of high-real-time scenarios such as brain-computer interfaces.

[0018] Preferably, in step (1), the collected neural or physiological signals are divided into segments of equal length according to a fixed time window.

[0019] Preferably, in step (1), each window is either overlapping or non-overlapping, providing a standardized input tensor for subsequent end-to-end model processing.

[0020] Preferably, in step (1), the use of a differentiable band-stop filter module to suppress power frequency and specific frequency band interference is as follows: a trainable symmetric impulse response filter kernel is integrated in a one-dimensional convolutional layer to adaptively suppress interference of specified frequency bands and their harmonic components in the signal.

[0021] Preferably, in step (1), the filter kernel of the band-stop filter is a symmetric sinc kernel, a symmetric Gaussian kernel, or a parameterized window function. The symmetry ensures that the convolution operation is equivalent to the traditional filter, thereby achieving frequency-domain selective suppression. The key parameters of the filter kernel include the center frequency, bandwidth, and window function parameters. Each key parameter is set as a learnable variable and is automatically optimized through the backpropagation mechanism of the deep learning model to achieve adaptive adjustment for different signal characteristics and application scenarios.

[0022] Preferably, in step (1), the adaptive extraction of the target frequency band through the bandpass filter layer with a learnable kernel function is as follows: the target frequency band components related to the analysis task are adaptively extracted from the original neural or physiological signals, and a trainable symmetric impulse response filter kernel is introduced into the one-dimensional convolutional layer; the bandpass filter amplifies the target frequency band and attenuates the irrelevant frequency band, and the center frequency and bandwidth of the filter kernel are dynamically adjusted as learnable variables during the training process.

[0023] Preferably, in step (2), time-frequency domain features are extracted from the preprocessed neural or physiological signals to achieve a fine representation of the transient changes and local frequency components of the signals; within the framework of the deep learning model, a trainable time-frequency analysis kernel is embedded in a one-dimensional convolutional layer to support complex parameterization to extract amplitude and phase information simultaneously; through end-to-end training, the center frequency, bandwidth, and shape parameters of the time-frequency analysis kernel are automatically optimized.

[0024] Preferably, in step (3), the output feature sequence is input into multiple convolutional channels, and each channel uses a different convolutional kernel size or receptive field to extract features from the signal; each convolutional channel works in parallel to capture multi-scale information of the signal.

[0025] Preferably, in step (3), the multi-scale parallel convolutional layer fusion includes trainable one-dimensional convolutional kernels with different lengths, strides, and number of channels. The multi-scale features after the parallel convolutional layer are merged into a unified feature vector by splicing or weighted summation. This feature vector is then input into the classification network for output. The classification network includes one or more fully connected layers and a neural network structure module that combines pooling, normalization, and activation functions. The final output is a multi-class or binary classification probability for subsequent application discrimination tasks.

[0026] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a neural and physiological signal processing device for differentiable filtering and wavelet analysis. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes: The signal segmentation module divides the acquired neural or physiological signals into segments of equal length according to a fixed time window. Differentiable band-stop filter module integrates a trainable symmetric impulse response filter kernel in a one-dimensional convolutional layer for adaptive suppression of interference in a specified frequency band and its harmonic components in a signal; The bandpass filtering module adaptively extracts target frequency band components relevant to the analysis task from the raw neural or physiological signals, and introduces a trainable symmetric impulse response filtering kernel into the one-dimensional convolutional layer; The time-frequency analysis module extracts time-frequency domain features from preprocessed neural or physiological signals, enabling precise characterization of transient changes and local frequency components of the signal. Multi-scale parallel convolutional layers and classification networks input the output feature sequences into multiple convolutional channels, each channel using a different convolutional kernel size or receptive field to extract features from the signal; each convolutional channel works in parallel to capture multi-scale information of the signal.

[0027] This invention relies on a unified computing architecture to integrate preprocessing, feature extraction, and classification modules into a deep learning framework. It accelerates computation through native operators and supports automatic optimization of backpropagation of filter parameters and wavelet basis functions, significantly improving the automation and real-time performance of neural and physiological signal processing. It is suitable for various dynamic application scenarios such as brain-computer interfaces and disease monitoring.

[0028] This invention uses band-stop filtering, band-pass filtering, and parameterized wavelet (time-frequency) analysis kernels as a basis, based on symmetric impulse response functions, and embeds them into a deep learning model in the form of differentiable and trainable parameters. This enables all frequency band parameters (such as center frequency, bandwidth, window function parameters, wavelet parameters, etc.) to be learned variables of the neural network. Through end-to-end training, the filtering and feature extraction operators are dynamically and adaptively optimized, and it is compatible with various implementation forms such as sinc kernel, symmetric Gaussian kernel, and Morlet wavelet.

[0029] This invention employs a multi-scale parallel convolutional structure to fuse time-domain, frequency-domain, and time-frequency features, integrating signal preprocessing, feature extraction, and classification into a unified deep learning framework. This architecture is compatible with hardware acceleration platforms such as GPUs, enabling end-to-end inference and training, effectively reducing overall system latency and meeting the engineering requirements for real-time analysis. The solution is widely applicable to the rapid analysis and classification of various physiological and neural signals, including but not limited to EEG, ECG, EMG, and MEG.

[0030] The symmetric design of the notch and bandpass kernels can be replaced with other symmetric kernel functions that satisfy frequency domain characteristics, such as cosine modulation kernels and Gaussian window kernels, adjusting the kernel shape without abandoning the cross-correlation convolution implementation. Similarly, the Morlet wavelet kernel in the time-frequency analysis module can be replaced with other symmetric wavelet basis functions (such as Gaussian derivative wavelets). The classification network can be replaced with lightweight models (such as EEGNet) or time-series models (such as Long Short-Term Memory Neural Networks LSTM) to adapt to different real-time or resource-constrained scenarios, depending on application requirements.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A neural and physiological signal processing method based on differentiable filtering and wavelet analysis, characterized in that: It includes the following steps: (1) The input signal is windowed, and the power frequency and specific frequency band interference are suppressed by the differentiable band-stop filter module. The target frequency band is adaptively extracted by the bandpass filter layer with learnable kernel function to achieve noise suppression and signal enhancement. (2) Input the preprocessed signal into the time-frequency analysis module, use the parameterized wavelet transform layer to generate the optimal wavelet kernel in real time, and combine the complex wavelet to retain the phase information to achieve adaptive optimization of time-frequency resolution and feature representation; (3) Features are fused through multi-scale parallel convolutional layers and the results are output using a lightweight classification network.

2. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 1, characterized in that: In step (1), the collected neural or physiological signals are divided into segments of equal length according to a fixed time window.

3. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 2, characterized in that: In step (1), each window may or may not overlap, providing a standardized input tensor for subsequent end-to-end model processing.

4. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 3, characterized in that: In step (1), the use of a differentiable band-stop filter module to suppress power frequency and specific frequency band interference is as follows: a trainable symmetric impulse response filter kernel is integrated in a one-dimensional convolutional layer to adaptively suppress the interference of specified frequency bands and their harmonic components in the signal.

5. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 4, characterized in that: In step (1), the filter kernel of the band-stop filter is a symmetric sinc kernel, a symmetric Gaussian kernel, or a parameterized window function. The symmetry ensures that the convolution operation is equivalent to the traditional filter, thereby achieving frequency-domain selective suppression. The key parameters of the filter kernel include the center frequency, bandwidth, and window function parameters. Each key parameter is set as a learnable variable and is automatically optimized through the backpropagation mechanism of the deep learning model to achieve adaptive adjustment for different signal characteristics and application scenarios.

6. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 5, characterized in that: In step (1), the target frequency band is adaptively extracted by the bandpass filter layer with a learnable kernel function as follows: the target frequency band components related to the analysis task are adaptively extracted from the original neural or physiological signals, and a trainable symmetric impulse response filter kernel is introduced into the one-dimensional convolutional layer; the bandpass filter amplifies the target frequency band and attenuates the irrelevant frequency band, and the center frequency and bandwidth of the filter kernel are dynamically adjusted as learnable variables during the training process.

7. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 6, characterized in that: In step (2), time-frequency domain features are extracted from the preprocessed neural or physiological signals to achieve a fine representation of the transient changes and local frequency components of the signals; within the deep learning model framework, a trainable time-frequency analysis kernel is embedded in a one-dimensional convolutional layer to support complex parameterization to extract amplitude and phase information simultaneously; through end-to-end training, the center frequency, bandwidth, and shape parameters of the time-frequency analysis kernel are automatically optimized.

8. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 7, characterized in that: In step (3), the output feature sequence is input into multiple convolution channels respectively. Each channel uses a different convolution kernel size or receptive field to extract features from the signal respectively. Each convolution channel works in parallel to capture multi-scale information of the signal.

9. The neural and physiological signal processing method based on differentiable filtering and wavelet analysis according to claim 8, characterized in that: In step (3), the multi-scale parallel convolutional layer fusion includes trainable one-dimensional convolutional kernels with different lengths, strides, and number of channels. The multi-scale features after the parallel convolutional layer are merged into a unified feature vector by splicing or weighted summation. This feature vector is then input into the classification network for output. The classification network includes one or more fully connected layers and a neural network structure module that combines pooling, normalization, and activation functions. The final output is a multi-class or binary classification probability for subsequent application discrimination tasks.

10. A neural and physiological signal processing device based on differentiable filtering and wavelet analysis, characterized in that: It includes: The signal segmentation module divides the acquired neural or physiological signals into segments of equal length according to a fixed time window. Differentiable band-stop filter module integrates a trainable symmetric impulse response filter kernel in a one-dimensional convolutional layer for adaptive suppression of interference in a specified frequency band and its harmonic components in a signal; The bandpass filtering module adaptively extracts target frequency band components relevant to the analysis task from the raw neural or physiological signals, and introduces a trainable symmetric impulse response filtering kernel into the one-dimensional convolutional layer; The time-frequency analysis module extracts time-frequency domain features from preprocessed neural or physiological signals, enabling precise characterization of transient changes and local frequency components of the signal. Multi-scale parallel convolutional layers and classification networks input the output feature sequences into multiple convolutional channels, each channel using a different convolutional kernel size or receptive field to extract features from the signal; each convolutional channel works in parallel to capture multi-scale information of the signal.

Images