CN-121971108-A - Lightweight signal classification method and device based on electroencephalogram signals

CN121971108ACN 121971108 ACN121971108 ACN 121971108ACN-121971108-A

Abstract

The invention discloses a lightweight signal classification method and device based on an electroencephalogram signal, and relates to the technical field of brain images, wherein the method comprises the steps of collecting the electroencephalogram signal, preprocessing, inputting the preprocessed signal into an improved depth residual error shrinkage network for feature extraction, integrating a blueprint separable convolution layer and a multi-directional collaborative attention mechanism into the network, and finally realizing signal classification based on the extracted features; the method has the advantages that the accuracy, the efficiency and the reliability of PD detection are remarkably improved in a basic clinical scene, and a feasible support is provided for real-time screening in a low-calculation-force environment.

Inventors

TANG JINGHONG
WANG CHUHAN
WU ZHIJING
GUO HANFEI
Zou Yayu
JU WEI

Assignees

四川大学

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. The lightweight signal classification method based on the electroencephalogram signals is characterized by comprising the following steps of: collecting an electroencephalogram signal and preprocessing the electroencephalogram signal; Inputting the preprocessed electroencephalogram signals into a pre-established improved depth residual error contraction network for feature extraction, wherein the improved depth residual error contraction network comprises at least one residual error block, and each residual error block comprises a blueprint separable convolution layer and a multi-directional collaborative attention mechanism; and classifying the electroencephalogram signals based on the extracted features to obtain classification results.
2. The method for classifying signals according to claim 1, wherein the blueprint separable convolution layer comprises: the method comprises the steps of performing channel projection and compression on input features to generate subspace feature representations; the method comprises the steps of mixing channels of the subspace characteristic representations to generate an intermediate characteristic diagram; And the method is used for carrying out deep convolution operation on the intermediate feature map, extracting space structure information and outputting a final feature map.
3. The method for classifying lightweight signals based on electroencephalogram signals according to claim 2, wherein said expression for channel projection and compression of input features comprises: Wherein, the Then the original signal matrix is represented Middle (f) A map of the characteristics of the individual input channels, An input original signal matrix; For the original signal matrix Is a subspace channel number of (2); to input channel Projection to subspace channel Is used for the linear weight of the (c), Representation of subspaces compressed for channels Middle (f) A characteristic map of the individual channels is provided, Is a subspace representation of the channel compression.
4. A lightweight signal classification method based on electroencephalogram signals as recited in claim 3, wherein said expression for channel blending said subspace feature representation comprises: Wherein, the To channel subspace Mix to the output channel Is used for the linear weight of the (c), In order to output the number of channels, Is a mixed intermediate characteristic diagram Middle (f) A map of the characteristics of the individual output channels, Is the intermediate feature map after mixing.
5. The method of claim 1, wherein the multi-directional collaborative attention mechanism comprises: the upper branch processing, respectively carrying out average pooling along the X direction and the Y direction to generate a direction attention diagram; The lower branch processing, respectively carrying out average pooling along the X direction and the Y direction, and generating a global direction attention map based on global statistics; and (3) fusion processing, namely fusing the outputs of the upper branch and the lower branch with the main characteristic to generate a point-by-point non-negative threshold value.
6. The method for classifying lightweight signals based on electroencephalogram signals according to claim 5, wherein the expression of the upper branch processing includes: the expression of the lower branch processing includes: The expression of the fusion process includes: In the formula, Is output by a trunk, and has the dimension of , 、、 Are all constant, wherein Is the number of channels to be processed, Is the height of the space which is the same as the height of the space, Is the space width; Is that Is positioned at Scalar value of position, in subsequent calculation to Is that Direction to Is that Direction of , For indexing the channel dimensions of the feature map, For indexing spatial height dimensions, expressed in spatial width Two-dimensional characteristic diagram obtained by carrying out average pooling on dimension , For indexing the channel dimensions of the feature map, For indexing spatial width dimensions, expressed in spatial height Carrying out average pooling on the dimension to obtain a two-dimensional feature map; Representing that all are A complete two-dimensional feature map of the composition, Representing that all are A complete two-dimensional feature map is formed; representation of progress The point convolution operation is performed with the result that, Indicating that a batch normalization operation is to be performed, Representing a sigmoid activation function; Representing edges Channel bottleneck coding of the direction; Is a rim Channel bottleneck coding of the direction; Is a rim A directional door of the direction; is a directional gate along the Y direction; Is that Is positioned in The scalar value of the location is used, Is a two-dimensional direction attention drawing obtained by the outer product of a two-dimensional door; Respectively the edges And Global gates of directions; Is a global gating signal In position Scalar values on; Is a global gating signal In position Scalar values on; Is that Is positioned in The scalar value of the location is used, A two-dimensional global attention graph obtained for global gate outer product; For the multiplication on an element-by-element basis, Two paths of weighting characteristics are respectively an upper branch and a lower branch, Is the fused intermediate feature; Is a fused representation; Representing the function of the ReLU activation, The final point-by-point non-negative threshold is butted with the point-by-point soft threshold shrinkage step of the trunk, The Softplus functions are represented.
7. The method of claim 1, wherein the operation procedure of each residual block in the improved depth residual contraction network comprises: extracting initial features by the first layer blueprint separable convolution layer; generating a point-by-point soft threshold by utilizing a multidirectional cooperative attention mechanism; Performing a point-wise soft threshold shrink operation to suppress noise; feature integration is carried out on the separable convolution layer through the second blueprint; the processed features are added to the input features by residual connection.
8. The method for classifying lightweight signals based on electroencephalogram signals according to claim 1, wherein the preprocessing comprises: Filtering the electroencephalogram signals by using a band-pass filter, and reserving a target frequency band to obtain first electroencephalogram signals; Removing eye movement and myoelectric artifacts in the first electroencephalogram signal by an independent component analysis method to obtain a second electroencephalogram signal; and carrying out standardization processing on the second electroencephalogram signals.
9. The method of claim 1, wherein the training process of the improved depth residual contraction network uses a composite objective function of weighted cross entropy loss and orthogonal regularization loss, expressed as: In the formula, In order to account for the total loss, In order to classify the cross entropy loss, Is an orthogonal regularization term that is used to determine, , And Are super parameters.
10. A lightweight signal classification device based on electroencephalogram signals, the device comprising: the preprocessing module is used for acquiring the brain electrical signals and preprocessing the brain electrical signals; The device comprises a feature extraction module, a feature extraction module and a feature extraction module, wherein the feature extraction module is used for inputting the preprocessed electroencephalogram signals into a pre-established improved depth residual error contraction network for feature extraction, the improved depth residual error contraction network comprises at least one residual error block, and each residual error block comprises a blueprint separable convolution layer and a multi-direction collaborative attention mechanism; And the classification module is used for classifying the electroencephalogram signals based on the extracted features to obtain classification results.

Description

Lightweight signal classification method and device based on electroencephalogram signals Technical Field The invention relates to the technical field of brain images, in particular to a lightweight signal classification method and device based on brain electrical signals. Background With the continuous rise of the incidence of neurodegenerative diseases, early screening of parkinson's disease is becoming an important research direction in the medical health field. The traditional diagnosis method relies on clinical symptom observation and medical image examination, and has the limitations of strong subjectivity, high equipment requirement, high cost and the like. The brain electrical signal is used as an important biomarker for reflecting the neural activity, and provides a new technical path for early detection of the Parkinson's disease due to the characteristics of noninvasive, real-time and low cost. In the related technology, the existing detection method based on the EEG signals has the problems that the traditional deep learning network has large parameters and is difficult to deploy in the limited computing power environment of the basic medical scene, the clinically collected EEG signals are easy to be interfered by the environment to influence the stability of feature extraction, and the model adaptability is reduced due to the distribution difference of cross-equipment and cross-crowd. Based on this, how to overcome the signal noise problem caused by the complex factors of environment, equipment and the like is needed to be solved because of the inefficiency of the model existing in the traditional EEG-based parkinsonism detection. Disclosure of Invention In view of the above problems, the invention provides a lightweight signal classification method and device based on electroencephalogram signals, which can effectively improve a model in a complex basic-level clinical scene and provide feasible technical support for the screening efficiency, accuracy and reliability of parkinsonism. In a first aspect, an embodiment of the present invention provides a lightweight signal classification method based on an electroencephalogram signal, including: collecting an electroencephalogram signal and preprocessing the electroencephalogram signal; Inputting the preprocessed electroencephalogram signals into a pre-established improved depth residual error contraction network for feature extraction, wherein the improved depth residual error contraction network comprises at least one residual error block, and each residual error block comprises a blueprint separable convolution layer and a multi-directional collaborative attention mechanism; and classifying the electroencephalogram signals based on the extracted features to obtain classification results. In some embodiments the blueprint separable convolution layer comprises: the method comprises the steps of performing channel projection and compression on input features to generate subspace feature representations; the method comprises the steps of mixing channels of the subspace characteristic representations to generate an intermediate characteristic diagram; And the method is used for carrying out deep convolution operation on the intermediate feature map, extracting space structure information and outputting a final feature map. In some embodiments, the expression for channel projection and compression of input features includes: Wherein, the Then the original signal matrix is representedMiddle (f)A map of the characteristics of the individual input channels,An input original signal matrix; For the original signal matrix Is a number of subspace channels.To input channelProjection to subspace channelIs used for the linear weight of the (c),Representation of subspaces compressed for channelsMiddle (f)A characteristic map of the individual channels is provided,Is a subspace representation of the channel compression. In some embodiments, the expression for channel projection and compression of input features includes: Wherein, the To channel subspaceMix to the output channelIs used for the linear weight of the (c),In order to output the number of channels,Is a mixed intermediate characteristic diagramMiddle (f)A map of the characteristics of the individual output channels,Is the intermediate feature map after mixing. In some embodiments, the multi-directional collaborative attention mechanism comprises: the upper branch processing, respectively carrying out average pooling along the X direction and the Y direction to generate a direction attention diagram; The lower branch processing, respectively carrying out average pooling along the X direction and the Y direction, and generating a global direction attention map based on global statistics; and (3) fusion processing, namely fusing the outputs of the upper branch and the lower branch with the main characteristic to generate a point-by-point non-negative threshold value. In some embodiments, the expression