Olfactory disorder electroencephalogram signal recognition method, system, device and storage medium

Abstract

This invention belongs to the field of biomedical signal processing technology, and relates to a method, system, device, and storage medium for recognizing electroencephalogram (EEG) signals in olfactory disorders. The method includes: S1: acquiring the raw multi-channel EEG signals to be processed, and preprocessing and slicing them to form an EEG slice sequence; S2: constructing a dual-path fusion model and pre-training the dual-path fusion model, inputting the EEG slice sequence into the pre-trained dual-path fusion model to extract fusion reconstruction features; S3: inputting the fusion reconstruction features into an improved KAN-GRU classification model after supervised classification training to generate the classification and recognition results of the EEG signals. This invention acquires the dynamic temporal features and static physiological features of EEG signals through a dual-path parallel architecture, effectively eliminating individual differences and phase delay interference. Combined with the strong nonlinear expression capability of the improved KAN-GRU, it significantly improves the accuracy and robustness of olfactory disorder classification.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing and neural engineering technology, and relates to a method, system, device and storage medium for recognizing electroencephalogram (EEG) signals of olfactory disorders, and in particular a method, system, device and storage medium for recognizing EEG signals of olfactory disorders based on adaptive phase calibration Mamba and multi-prototype learning. Background Technology

[0002] Olfactory dysfunction refers to the partial or complete loss of olfactory perception, including hyposmia (low sense of smell) and anosmia (loss of smell). Clinical studies have shown that olfactory dysfunction is not only a significant factor in decreased quality of life, but also an early biomarker for neurodegenerative diseases such as Parkinson's Disease (PD), Alzheimer's Disease (AD), and multiple sclerosis. Particularly during the COVID-19 pandemic, acute olfactory dysfunction was identified as a key diagnostic indicator of viral infection. Therefore, developing objective, accurate, and reproducible methods for assessing olfactory function has significant clinical value and social implications.

[0003] Currently, commonly used clinical methods for olfactory assessment mainly fall into two categories: psychophysical tests and neurophysiological tests. Psychophysical tests, such as the Sniffin' Sticks test, the T&T Olfactometer test, and the University of Pennsylvania Olfactory Identification Test (UPSIT), rely on the subject's subjective description and active cooperation, requiring the subject to possess normal cognitive function and language expression ability. For subjects with cognitive impairment, impaired consciousness, excessively young age, or a tendency to feign illness, these methods are difficult to obtain reliable results. Furthermore, the results of psychophysical tests are easily affected by environmental odors, the subject's emotional state, and the operator's procedural compliance, resulting in poor repeatability and objectivity.

[0004] Electroencephalography (EEG), a non-invasive neurophysiological detection technique, can objectively reflect the integrity and functional state of the olfactory neural pathway by recording the synchronized activity of postsynaptic potentials in groups of neurons in the cerebral cortex. When olfactory stimuli (such as odor molecules) act on olfactory receptor neurons in the olfactory epithelium, the generated neural signals are transmitted through the olfactory bulb and olfactory tract to the primary olfactory cortex (piriform cortex) and secondary olfactory-related areas (orbitofrontal cortex, amygdala, etc.), forming detectable evoked potentials. Therefore, analyzing the temporal and frequency characteristics, spatial distribution patterns, and dynamic evolution of olfactory event-related potentials (OERPs) provides a physiological basis for the objective assessment of olfactory disorders. However, existing EEG identification methods for olfactory disorders suffer from problems such as insufficient temporal modeling ability, poor adaptability to individual physiological differences, phase delay interference, and limited classifier expressive ability. Against this backdrop, there is an urgent need for a novel method, system, device, and storage medium for recognizing electroencephalogram (EEG) signals in olfactory disorders. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method, system, device, and storage medium for recognizing electroencephalogram (EEG) signals of olfactory disorders. It is an EEG recognition method for olfactory disorders based on adaptive phase calibration Mamba and multi-prototype learning. By using a dual-path parallel architecture to acquire the dynamic temporal features and static physiological features of EEG signals, it effectively eliminates individual differences and phase delay interference. Combined with the strong nonlinear expression capability of the improved KAN-GRU, it significantly improves the accuracy and robustness of olfactory disorder classification.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for recognizing electroencephalogram (EEG) signals in cases of olfactory dysfunction, characterized by comprising the following steps: S1: Acquire the raw multi-channel EEG signal to be processed, and preprocess and slice the EEG signal to form an EEG slice sequence; S2: Construct a dual-path fusion model and pre-train it. Input the EEG slice sequence into the pre-trained dual-path fusion model to extract fusion reconstruction features. The dual-path fusion model includes an associated topology calibration Mamba model, a multi-prototype Riemann codebook model, and a feature fusion module. The associated topology calibration Mamba model is used to perform temporal modeling through selective state space scanning to extract dynamic temporal features rich in olfactory stimulus response patterns from the EEG slices, and uses the associated topology calibration mechanism to perform phase alignment of the dynamic temporal features to obtain temporal features. The multi-prototype Riemann codebook model is used to calculate the covariance matrix of the EEG slices to construct positive definite symmetric manifold nodes, and then project them onto the tangent space and match and weightedly fuse them with the Riemann codebook of learnable Riemann prototypes to obtain prototype features representing individual physiological background. The feature fusion module is used to weightedly fuse the temporal features and prototype features to form fusion reconstruction features. S3: The fused reconstructed features are input into the improved KAN-GRU classification model after supervised classification training to generate the classification and recognition results of the EEG signals. The improved KAN-GRU classification model includes a KAN-GRU module, an attention aggregation module, and a classification module. The KAN-GRU module is a deep recurrent temporal encoder composed of multiple KAN-GRU units stacked along the time dimension, used to extract the hidden state of each EEG slice based on the fused reconstructed features. The attention aggregation module is used to perform weighted aggregation of the hidden states of all EEG slices in the EEG slice sequence through an attention mechanism to obtain global features. The classification module is used to obtain the classification and recognition results of the EEG signals based on the global features.

[0007] Preferably, the associated topology calibration Mamba model specifically includes: The selective state-space scanning module is used to recursively scan EEG slices along the channel dimension, in order to To update the formula, the hidden state matrix is ​​iteratively updated using the hidden state matrix from the previous time step and the input data of the current EEG slice. After iterating through the entire EEG slice, the overall hidden state matrix is ​​output. ,in, It is the first The hidden state matrix at time step 1. It is a learnable state transition matrix. It is a learnable input projection matrix. The first EEG slice Input data at any time, It is the first The hidden state matrix at time step 1. The initial hidden state matrix is ​​initialized to an all-zero matrix. A relational topology calibration module, used to calibrate the two-dimensional hidden state matrix. Remodeled into a 3D hidden state matrix ; using the reshaped 3D hidden state matrix A feature association graph is constructed based on the feature association graph. The vertex set of the feature association graph corresponds to all EEG channels, and the edge set represents the functional coupling relationship between channels. Based on the feature association graph, the dynamic adjacency matrix between channels is calculated. Based on the dynamic adjacency matrix, the two-dimensional offset for phase calibration is predicted. Based on the two-dimensional offset, bilinear interpolation is used to correct the position and align the phase of the reconstructed three-dimensional hidden state matrix to obtain the aligned hidden state matrix. The aligned hidden state matrix is ​​restored to two dimensions and mapped through the output projection matrix to obtain the temporal features.

[0008] Preferably, the features of the multi-prototype Riemann codebook model specifically include: The tangent space projection and codebook matching module is used to calculate the spatial covariance matrix of EEG slices, and construct positive definite symmetric manifold nodes representing the topological coupling characteristics of neural activity based on the spatial covariance matrix. The positive definite symmetric manifold nodes are projected onto the corresponding Euclidean tangent space using matrix logarithm operations to obtain linearized Riemann representation vectors. The Euclidean distance between the Riemann representation vectors and the prototype vectors in the preset learnable Riemann prototype codebook is calculated. The most similar vector is selected based on the Euclidean distance. A learnable Riemannian prototype; the EEG slice pairs are calculated using a similarity function. Normalized membership weights of a learnable Riemannian prototype; The prototype weighted fusion and geometric restoration module is used to apply the normalized membership weights to the selected... We perform weighted summation and exponentiation on the prototype vectors of learnable Riemann prototypes; then we flatten the features and transform the dimensions of the weighted summation and exponentiation results to obtain the prototype features.

[0009] Preferably, during the pre-training of the dual-path fusion model, the model parameters of the dual-path fusion model are jointly optimized based on reconstruction loss and quantization loss, and the loss function during joint optimization is... for:

[0010]

[0011]

[0012] in, For reconstruction loss; For the first Fusion and reconstruction features of individual EEG slices; For the first Original features of an EEG slice; To quantify the loss; This is the balance coefficient; Indicates the first Riemann representation vectors of computer slices; Indicates the first The prototype features corresponding to each computer slice; This indicates that the gradient operation has been stopped. This represents the Frobenius norm.

[0013] Preferably, in step S3, the KAN-GRU module extracts the hidden state of each EEG slice based on the fused reconstruction features as follows: S31: Each of the KAN-GRU units uses a linear transformation of the standard GRU plus a Sigmoid activation to calculate the gate control signal for resetting the gate. and update the door's gate control signal :

[0014]

[0015] in, This represents vector concatenation. The weight matrix is ​​a learnable matrix. For bias, Use the Sigmoid activation function; For the first The characteristics of time-matter fusion and reconstruction; For the first The hidden state at all times The initial hidden state is initialized as an all-zero matrix; S32: Each of the KAN-GRU units will... Moment-by-moment fusion and reconstruction features After being modulated by the reset gate, the first Hidden state of time The vectors are concatenated to obtain the input vector of the KAN layer. :

[0016] Where ⊙ represents element-wise multiplication; S33: The KAN layer of each KAN-GRU unit is based on the input vector. The first is obtained by combining learnable univariate spline functions. Hidden state of time ; S34: The hidden state output by the last KAN-GRU unit at the last moment. This is the hidden state of the EEG slice.

[0017] Preferably, in step S3, the attention aggregation module performs weighted aggregation of the hidden states of all EEG slices in the EEG slice sequence through an attention mechanism to obtain global features, specifically as follows:

[0018] in, As a global feature, The total number of EEG slices in the EEG slice sequence. It is the first The hidden state of a computer slice. These are learnable attention weight parameters. This is the index used during aggregation.

[0019] Preferably, in step S3, the classification module obtains the classification and recognition result of the EEG signal based on the global features by: classifying the global features... The input is a fully connected layer, which is activated via Softmax and outputs the probability distribution of the subjects belonging to each category. :

[0020] in, The classification weight matrix is ​​a learnable matrix. This is the bias vector.

[0021] Furthermore, the present invention also provides a brainwave signal recognition system for olfactory disorders, characterized in that it comprises: The EEG signal acquisition and processing module is used to acquire the raw multi-channel EEG signals to be processed, and to preprocess and slice the EEG signals to form an EEG slice sequence. A fusion reconstruction feature acquisition module is used to construct a dual-path fusion model and pre-train the model. The EEG slice sequence is input into the pre-trained model to extract fusion reconstruction features. The dual-path fusion model includes an associated topology calibration Mamba model, a multi-prototype Riemann codebook model, and a feature fusion module. The associated topology calibration Mamba model performs temporal modeling through selective state space scanning to extract dynamic temporal features rich in olfactory stimulus response patterns from the EEG slices, and uses an associated topology calibration mechanism to phase-align these dynamic temporal features to obtain temporal features. The multi-prototype Riemann codebook model calculates the covariance matrix of the EEG slices to construct positive definite symmetric manifold nodes, projects them onto the tangent space, and then matches and weights them with a Riemann codebook of learnable Riemann prototypes to obtain prototype features representing individual physiological backgrounds. The feature fusion module performs weighted fusion of the temporal features and prototype features to form fusion reconstruction features. An EEG signal classification and recognition module is used to input the fused reconstructed features into an improved KAN-GRU classification model after supervised classification training to generate EEG signal classification and recognition results. The improved KAN-GRU classification model includes a KAN-GRU module, an attention aggregation module, and a classification module. The KAN-GRU module is a deep recurrent temporal encoder composed of multiple KAN-GRU units stacked along the time dimension, used to extract the hidden state of each EEG slice based on the fused reconstructed features. The attention aggregation module is used to perform weighted aggregation of the hidden states of all EEG slices in the EEG slice sequence through an attention mechanism to obtain global features. The classification module is used to obtain the EEG signal classification and recognition results based on the global features.

[0022] Furthermore, the present invention also provides a device for recognizing electroencephalogram (EEG) signals of olfactory disorders, characterized in that it comprises: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the olfactory disorder EEG signal recognition method as described above. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements the steps of the olfactory disorder EEG signal recognition method as described above.

[0023] Compared with the prior art, the olfactory disorder EEG signal recognition method, system, device and storage medium of the present invention have one or more of the following beneficial technical effects: (1) By using the RTC-Mamba model, the selective state space scanning mechanism is used to perform temporal modeling of content awareness, and the phase delay calibration is performed by combining the correlation graph convolutional network. This effectively solves the problem of feature misalignment caused by individual differences and neural conduction delay in EEG signals, improves the robustness of temporal features, and achieves accurate temporal modeling and calibration.

[0024] (2) By using a multi-prototype Riemann codebook model, the dynamic temporal features of EEG signals are decoupled from the static prototype features that represent the individual’s inherent physiological background. Stable and interpretable physiological state representations are extracted using Riemann geometry and codebook mechanisms, which enhances the adaptability to physiological differences among different subjects and achieves physiological background decoupling and robustness enhancement.

[0025] (3) By jointly pre-training the reconstruction loss and quantization loss, the dual-path fusion model can be optimized in a coordinated manner, ensuring the integrity and discriminativeness of the extracted features, and providing high-quality input features for subsequent classification tasks.

[0026] (4) In the classification stage, the improved KAN-GRU model is used. The nonlinear expression ability of the KAN layer is used to replace the linear transformation of the traditional GRU, which can more effectively capture the complex nonlinear dynamic patterns in the EEG signal, thereby improving the accuracy of olfactory disorder classification. Attached Figure Description

[0027] Figure 1 This is a flowchart of the electroencephalogram (EEG) signal recognition method for olfactory disorders according to the present invention.

[0028] Figure 2 This is a schematic diagram of the dual-path fusion model of the present invention.

[0029] Figure 3 This is a schematic diagram of the associated topology calibration Mamba model of the dual-path fusion model of the present invention.

[0030] Figure 4 This is a schematic diagram of the multi-prototype Riemann codebook model of the dual-path fusion model of the present invention.

[0031] Figure 5 This is a schematic diagram of the improved KAN-GRU classification model of the present invention.

[0032] Figure 6 This is a schematic diagram of the olfactory disorder EEG signal recognition system of the present invention.

[0033] Figure 7 This is a structural block diagram of the electroencephalogram (EEG) signal recognition device for olfactory disorders of the present invention. Detailed Implementation

[0034] Before detailing any embodiment of the invention, it should be understood that the invention, in its application, is not limited to the details of the construction and arrangement of the components set forth in the following description or illustrated in the following figures. The invention can have other embodiments and can be practiced or carried out in various ways. Furthermore, it should be understood that the wording and terminology used herein are for descriptive purposes and should not be considered limiting. The use of “comprising” or “having” and variations thereof in this invention is intended to cover the items set forth below and their equivalents, as well as any additional items. Unless otherwise specified or limited, the terms “installation,” “connection,” “support,” and “linkage,” and variations thereof are used broadly and cover both direct and indirect installation, connection, support, and linking. Moreover, “connection” and “linkage” are not limited to physical or mechanical connections or links. Furthermore, firstly, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention. Secondly, the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple. The term "a" should not be construed as a limitation on the quantity.

[0035] Figure 1 A flowchart of the electroencephalogram (EEG) signal recognition method for olfactory disorders of the present invention is shown. Figure 1 As shown, the method for identifying electroencephalogram (EEG) signals of olfactory disorders according to the present invention includes the following steps: S1: Acquisition and processing of electroencephalogram (EEG) signals.

[0036] The raw multichannel EEG signals to be processed are acquired, and the EEG signals are preprocessed and sliced ​​to form an EEG slice sequence.

[0037] Specifically, acquiring raw multichannel EEG signals , For the number of channels, This represents the total signal duration. Typically, a 64-lead EEG cap conforming to the international 10-20 system standard is used to collect EEG signals from subjects with varying degrees of olfactory impairment. The sampling rate is set to 1000Hz, and the acquired raw multi-channel EEG signals are... Its dimensions are , This represents the total duration of the signal.

[0038] The acquired raw multichannel EEG signals Preprocessing was performed in three aspects, including but not limited to: First, baseline correction was performed by subtracting the mean of the EEG signal in each channel within 200 ms before stimulus presentation to eliminate DC drift; then, a bandpass filter from 0.5 Hz to 45 Hz was used to remove low-frequency drift and high-frequency electromyographic noise; finally, starting from the stimulus presentation time, the EEG signal 2 seconds after stimulus (i.e., 2000 sampling points) was selected as an experimental EEG signal.

[0039] After baseline correction and filtering, all EEG signals used in the experiment were sliced ​​along the time axis. If the total signal duration... To splice together multiple experiments, it is divided into A length of EEG slices For example, each EEG slice The dimension is , The value range is 1 to .

[0040] S2: Feature fusion and reconstruction.

[0041] A dual-path fusion model is constructed and pre-trained. The EEG slice sequence is then input into the pre-trained dual-path fusion model to extract fusion reconstruction features.

[0042] like Figure 2 As shown, the dual-path fusion model includes an associated topology calibration Mamba model, a multi-prototype Riemann codebook model, and a feature fusion module.

[0043] 1. Correlate topology to calibrate the Mamba model.

[0044] Electroencephalogram (EEG) signals exhibit long-range temporal dependence, with neural responses evoked by olfactory stimuli often evolving over hundreds of milliseconds to seconds after stimulus presentation. Traditional recurrent neural networks (RNNs) or Transformers suffer from gradient vanishing or excessive computational complexity when processing long sequences. To address this, this invention designs an RTC-Mamba (Relational Topological Calibration Mamba) encoding model. Here, R stands for Relational, indicating that the model uses an Aggregate Graph Convolutional Network (AGCN) to construct a feature correlation graph, quantifying the functional coupling strength between channels; T stands for Topological, indicating that the model calculates a dynamic adjacency matrix based on topological relationships and performs positional correction using a two-dimensional offset vector; C stands for Calibration, indicating that the model aligns the phase delays between different subjects or slices through a relational topological mechanism; and Mamba refers to the use of the Mamba model, employing a selective state-space scanning operator for content-aware temporal modeling. The associated topology calibration Mamba model is used to perform temporal modeling through selective state space scanning to extract dynamic temporal features rich in olfactory stimulus response patterns from the EEG slices, and to use the associated topology calibration mechanism to perform phase alignment of the dynamic temporal features to obtain temporal features.

[0045] like Figure 3 As shown, the associated topology calibration Mamba model specifically includes: (1) Selective state space scanning module.

[0046] The preprocessed first Each EEG slice is recorded as ,in, For the number of channels, This refers to the length of the EEG slice, i.e., the number of sampling points within the EEG slice. To accommodate the sequential processing mode of the selective state-space scanning module, the EEG slices are... Considered to be of length The time series, i.e. ,in, EEG slices In the Time, that is, the first All sampling points The signal value vector of the nth channel, this vector is the signal value vector of the nth channel during the scanning process. Step, that is, the first step The current input data at any given moment.

[0047] The following introduces three learnable parameter matrices for state-space modeling, denoted as: A: The state transition matrix determines the intrinsic evolution pattern of the hidden states; B: Input projection matrix, responsible for projecting the current input data. Mapped to the state space; C: Output projection matrix, responsible for decoding the required features from the hidden state.

[0048] The selective state-space scanning module performs a recursive scan along the channel dimension. At the start of the scan, the initial hidden state... It is initialized as a matrix of all zeros, i.e. This indicates that there is no prior state information before the first input data is received. From the first... From time t, the hidden state matrix The update rules are as follows:

[0049] in, It is the current moment of the EEG slice, that is, the first... Input data at any given time; It is the previous moment, that is, the first moment. The hidden state matrix at time step 1.

[0050] After iterating through the entire EEG slice sequence, the overall dynamic hidden state matrix is ​​output. We completed the preliminary encoding of EEG temporal features and state evolution modeling.

[0051] (2) Relationship topology calibration module.

[0052] There is a common problem of phase delay inconsistency between different subjects or different EEG slices. If the hidden state matrix output by the selective state space scanning module is directly... For subsequent feature fusion and reconstruction, biases in feature alignment will significantly affect model performance. Therefore, this invention introduces an association topology calibration mechanism, namely, a relationship topology calibration module, which calibrates the hidden state matrix... Nonlinear alignment is performed, and the first feature, namely the temporal feature, is obtained through output projection.

[0053] To facilitate the construction of the feature association graph, the hidden state matrix is ​​first processed. Perform a dimension reshaping operation, that is, a dimensional transformation, to convert it from the original two-dimensional form into a three-dimensional hidden state matrix. (satisfy That is, the timing characteristics of each channel become the shape of A two-dimensional matrix.

[0054] Then, using the reshaped 3D hidden state matrix Constructing a feature association graph based on Among them, vertex set Corresponding to all EEG channels, edge set This represents the functional coupling relationship between channels, and its connection weight is obtained from the subsequently calculated dynamic adjacency matrix. elements in Sure.

[0055] To quantify the functional coupling strength of different channels under the current olfactory stimulus, two learnable parameter matrices are introduced. and This is used to project the temporal characteristics of each channel into the query space and the key space. Specifically, for the first... Each channel has its corresponding feature matrix, i.e., the hidden state matrix. along Dimensions (feature dimensions) are respectively with and Multiply to obtain the query representation matrix. The bond representation matrix .

[0056]

[0057] Interchange the query representation and key representation of all channels along Average pooling is performed on the dimension (time segment dimension) to obtain the aggregate query vector for each channel. With aggregated bond vector .

[0058]

[0059] in, The projected aggregate query vector and aggregated bond vector The dimension; Indicates the first The first channel, the first The feature vectors corresponding to each time segment (dimension: ); Indicates the first The first channel, the first The query representation corresponding to each time segment; Indicates the first The first channel, the first The key representation corresponding to each time segment.

[0060] Based on the above aggregate query vector With aggregated bond vector Calculation channel With channel The attention weights between them are determined, and the dynamic adjacency matrix is ​​obtained by Softmax normalization. The specific formula is:

[0061] in, Indicates channel With channel Connection weights between them; It is the index used during normalization.

[0062] Obtaining the dynamic adjacency matrix Subsequently, the Association Graph Convolutional Network (AGCN) executes a message-passing mechanism, aggregating the neighborhood information of each channel into its own representation. To facilitate matrix operations, the reshaped 3D hidden state matrix is ​​first... Flattened into a two-dimensional matrix . Two-dimensional matrix Input linear layer, based on dynamic adjacency matrix Calculate the two-dimensional offset used for phase calibration. .

[0063]

[0064] in, Reshaped into Each channel In position The two-dimensional offset at that point is denoted as .in, It is the offset in the time segment dimension, used to correct the temporal phase deviation caused by neural conduction delay or individual reaction time differences; It is the offset in the feature dimension, used to correct feature space deviations caused by nonlinear coupling or spatial correlation differences between channels. Next, precise phase alignment of the hidden states is achieved through bilinear interpolation. A normalized two-dimensional coordinate grid is established. Among them, grid Corresponding to the The time segment and the first The original integer coordinates of the feature dimension. Each channel has its offset superimposed onto a reference grid to obtain the adaptive sampling coordinates for that channel. .

[0065]

[0066] because and The values ​​are usually non-integers and require the use of a hidden state matrix. Perform bilinear interpolation on the known integer values. For each non-integer coordinate... Take the coordinates of its four nearest integer neighbors to form a neighborhood set. The interpolated hidden state matrix In position value at It is given by the following formula:

[0067] This ensures the continuity and smoothness of the interpolation.

[0068] For all After performing the above operations on each channel, the aligned hyperdynamic feature tensor is obtained. Then to Perform a dimension reduction operation, that is, a dimensional transformation, to convert it from three dimensions. Convert to two dimensions ,satisfy The calibrated hidden state matrix is ​​obtained. .

[0069] Finally, mapping is performed, that is, mapping the calibrated hidden state matrix. The learnable output projection matrix mentioned above Multiplying them yields the first feature, namely, the temporal feature. .

[0070]

[0071] 2. Multi-prototype Riemann codebook model.

[0072] The multi-prototype Riemann codebook model is used to calculate the covariance matrix of EEG slices to construct positive definite symmetric manifold nodes, and then project them onto the tangent space for matching and weighted fusion with the Riemann codebook of learnable Riemann prototypes to obtain prototype features that characterize the individual's physiological background.

[0073] like Figure 4 As shown, the multi-prototype Riemann codebook model specifically includes: (1) Tangent space projection and codebook matching module.

[0074] The tangent space projection and codebook matching module aims to extract time-independent static spatial topological features representing an individual's inherent physiological background from EEG slices and perform codebook matching. The spatial covariance matrix of EEG signals can effectively reflect the synchronous coupling strength between different brain regions and is an important statistic for describing the functional connectivity state of the brain. However, the covariance matrix has positive definite symmetry and is located in a Riemannian manifold rather than Euclidean space; direct arithmetic operations on it will destroy its geometric structure.

[0075] In this invention, for the preprocessed first... Each EEG slice is recorded as ,middle, For the number of channels, This represents the number of sampling points within the EEG slice. First, the time-mean value is calculated independently for each channel to obtain the mean vector. :

[0076] in, No. The first EEG slice At the moment, that is, the moment A vector of signal values ​​at each time step.

[0077] Centering the EEG slices yields a zero-mean matrix. :

[0078] in, Indicates length is A vector of all 1s.

[0079] Based on this, the spatial covariance matrix between channels is calculated, and a minimal regularization term is added to ensure that the matrix is ​​strictly positive definite, thus constructing a positive definite symmetric manifold node that characterizes the spatial coupling properties of neural activity. :

[0080] in, For the smallest regularization term ( , (The identity matrix).

[0081] Since Riemannian manifolds do not possess the Euclidean addition property, this invention employs matrix logarithmic mapping to transform the nodes of positive definite symmetric manifolds. Projecting onto the tangent space of the node in the positive definite symmetric manifold yields a linearized representation while preserving the geometric structure. Specifically, this operation is performed on the nodes of the positive definite symmetric manifold. Perform standard eigenvalue decomposition:

[0082] in, for The One eigenvalue; The corresponding orthogonal eigenvector matrix; This represents a diagonal matrix with given elements. The eigenvalues ​​mentioned above... orthogonal eigenvector matrix Instead of learnable model parameters, they are derived from the positive definite symmetric manifold nodes of the currently input EEG slice. The eigenvalues ​​are determined directly by standard numerical eigenvalue decomposition algorithms. In practice, the above eigenvalue decomposition can be implemented by directly calling standard functions in numerical computing libraries (such as torch.linalg.eigh). Whenever a new EEG slice is input, the corresponding EEG slice needs to be... Perform eigenvalue decomposition in real time to obtain its complete feature spectrum and feature space structure.

[0083] Using matrix logarithm operations Calculate the nodes of a positive definite symmetric manifold Symmetric matrix representation in tangent space :

[0084] The operation procedure for matrix logarithm is as follows: First, perform the following operations... Performing the above eigenvalue decomposition yields and Then, take the natural logarithm of each eigenvalue element by element. Finally, the eigenvalues ​​after taking the logarithm are placed back into the diagonal matrix in their original positions and then rearranged into the original orthogonal eigenvector matrix. Multiplying the matrix by its transpose and rearranging them yields a symmetric matrix. In practice, this operation can be performed by directly calling the standard matrix logarithm function in the numerical computing library (such as torch.linalg.logm in PyTorch), which is essentially a combination of eigenvalue decomposition and element-wise scalar logarithm operation.

[0085] In this process, the logarithmic operation is applied element-wise to each eigenvalue. The resulting symmetric matrix Located in Euclidean space, it will serve as the first The Riemann representation vectors of each EEG slice will subsequently be used in matrix form to directly participate in the distance calculation and weighted fusion with each prototype node of the Riemann prototype codebook.

[0086] Due to physiological differences such as skull thickness, brain structure morphology, and neural conduction efficiency, the spatial covariance structure of different subjects exhibits significant individual variability. If such static background features are mixed with dynamic response features in the model, it will severely interfere with the model's extraction of olfactory stimulus-related patterns. Therefore, this invention introduces a multi-prototype mechanism, using a learnable codebook for soft matching to obtain stable and interpretable representations of individual physiological states.

[0087] Create a containing A Riemann codebook that can learn Riemann archetypes. Each learnable Riemann archetype This represents a physiological state. Calculate the Riemann representation vector of the current slice. Log-Euclidean distance between each learned Riemann prototype :

[0088] in, Denotes the Frobenius norm; This represents the matrix trace operation.

[0089] To achieve continuous representation of physiological states and enhance robustness, the nearest neighbor is selected. A learnable Riemannian archetype ( ), forming a set of neighboring prototypes .

[0090] Using temperature coefficient The Softmax similarity kernel function calculates the current slice for the... The first learnable Riemannian archetype Normalized membership weights of learnable Riemannian prototypes :

[0091] in, This is the index for normalization.

[0092] (2) Prototype weighted fusion and geometric restoration module.

[0093] The prototype weighted fusion and geometric restoration module is used to determine the normalized membership weights. For the selected A learnable Riemann prototype is subjected to weighted summation and exponentiation; the results of weighted summation and exponentiation are flattened and dimensionally transformed to obtain prototype features.

[0094] Specifically, based on the membership weights For the selected A learnable Riemannian prototype performs a linear weighted summation to generate prototype features that represent linear representations of the subject's inherent physiological background. :

[0095] To restore the geometric integrity of this physiological background feature in the EEG spatial coupling structure, it is necessary to transform the prototype feature represented linearly in the tangent space. The projection is then reprojected back to a positive definite symmetric manifold. This invention employs matrix exponential mapping. Implement this inverse transformation. Specifically, for the prototype features of the linear representation... Perform standard eigenvalue decomposition:

[0096] in, For eigenvalues; It is an orthogonal eigenvector matrix.

[0097] Then the SPD manifold node corresponding to this symmetric matrix It is given by the following formula:

[0098] Exponential mapping ensures It still satisfies the symmetric positive definiteness, that is, all its eigenvalues ​​are strictly greater than zero, thus making it effectively located on the original Riemannian manifold.

[0099] Then Flattening by row yields the flattened spatial topological features. :

[0100] To ensure that the output feature shape matches the input shape, this invention introduces a learnable linear decoding layer, which uses a learnable weight matrix. Flattened spatial topological features This is mapped to a high-dimensional space. Subsequently, a reshaping operation is performed. Finally, the second feature, namely the prototype feature, is obtained. :

[0101] 3. Feature fusion module.

[0102] The feature fusion module is used to perform weighted fusion of the temporal features and prototype features to form fused reconstructed features.

[0103] Specifically, learnable weight parameters are introduced. and The features of the two paths, namely, the time series features. and prototype features Weighted fusion is performed to obtain the fused and reconstructed features. .

[0104]

[0105] In this invention, after constructing the dual-path fusion model, it is jointly optimized and pre-trained using reconstruction loss and quantization loss. The reconstruction loss compares the differences between the original EEG slices and the reconstructed slices in the masked region, while simultaneously optimizing the associated topology calibration Mamba model and the multi-prototype Riemann codebook model to improve the temporal feature extraction capability and the accuracy of physiological background representation. The quantization loss is specifically used to update the multi-prototype Riemann codebook model, enhancing the stability of physiological state representation by calculating the alignment error between the Riemann representation vector and the weighted prototype features. Together, these two methods achieve collaborative training of the dual-path fusion model.

[0106] Specifically, during the input phase, the original EEG slice sequence is randomly masked (e.g., randomly occluding 15% of the time segment), and only the visible portions are fed into the dual-path fusion model. The fused and reconstructed features are then... According to the original EEG slice division method, corresponding to the The slice, denoted as the first slice. The reconstruction results of the EEG slices are as follows (Right now The corresponding number in the middle (part of an EEG slice), note the first The true signal of the original EEG slice is Reconstruction error is calculated only for the masked regions to drive the dual-path fusion model to infer the occluded EEG signals from the visible portion. Reconstruction loss. Defined as the sum of the mean square errors of the reconstructed results at all masked slice locations and the original signal:

[0107] To further enhance the stability and interpretability of physiological state representations in multi-prototype Riemann codebook models, vector quantization loss is introduced in addition to reconstruction loss. This is used to optimize the multi-prototype Riemann codebook model.

[0108] Record No. The Riemann representation vector of each EEG slice is Let the prototype feature of the linear representation corresponding to this slice be . Quantification of loss Defined as:

[0109] in, This indicates that the gradient operation is stopped (stop-gradient), meaning that the operand is treated as a constant during backpropagation and the gradient is not propagated. Denotes the Frobenius norm. and All are located in the tangent space Symmetric matrix; This is a balancing coefficient used to adjust the relative weights between the two losses.

[0110] In the above formula, the first term by To optimize variables, gradients flow only to learnable Riemann prototypes in the codebook. (Through membership weights) (Transmission), prompting the weighted combined prototype features to approximate the output Riemann representation vector, thereby updating the learnable Riemann prototype to better cluster discretized physiological states; the second term by To optimize variables and make the output Riemann representation vectors converge towards the fixed weighted prototype features, thereby enhancing the compactness and consistency of feature representations.

[0111] By employing this alternating stopping gradient strategy, bidirectional alignment of the multi-prototype Riemann codebook model is ultimately achieved, resulting in a stable and semantically interpretable discretized representation of physiological states.

[0112] Combining the two losses mentioned above, the total loss function of the dual-path fusion model during the pre-training stage is the sum of the reconstruction loss and the quantization loss:

[0113] During pre-training, the Adam optimizer is used to perform end-to-end joint parameter updates on the following model components via backpropagation: (1) Correlation topology calibration Mamba model: including learnable parameter matrix Query matrix Key matrix The linear layers in the graph convolutional network, and the output projection matrix. wait; (2) Multi-prototype Riemann codebook model: including Riemann codebooks with learnable Riemann prototypes. Linear decoding layer, etc.; (3) Feature fusion module: including a learnable fusion weight matrix and .

[0114] The pre-training process continues until the total loss is reached. Convergence. After training, freeze all model parameters and use the dual-path fusion model as a fixed feature extractor to provide high-quality fusion reconstruction features for the subsequent second-stage supervised classification prediction task based on the improved KAN-GRU. .

[0115] In practical use, the newly acquired EEG slice sequence is input into the trained dual-path fusion model to extract fusion reconstruction features for classification prediction. .

[0116] S3: Classification and recognition of electroencephalogram (EEG) signals.

[0117] The fused reconstructed features are input into an improved KAN-GRU classification model trained under supervised classification to generate classification and recognition results for EEG signals. The improved KAN-GRU classification model is used to capture complex nonlinear temporal dependencies from the fused reconstructed features and ultimately output a graded prediction result for olfactory disorders.

[0118] like Figure 5 As shown, the improved KAN-GRU classification model includes a KAN-GRU module, an attention aggregation module, and a classification module. In this invention, the overall computational flow of the improved KAN-GRU classification model is as follows: First, the gate control signals are calculated by the reset gate and the update gate respectively. Then, the current input is concatenated with the historical hidden state modulated by the reset gate and fed into the KAN layer to generate candidate hidden states. Finally, the current hidden state is obtained by convex combination of the historical state and the candidate state through the update gate. After stacking multiple such units to form a deep encoder, the hidden states of each time segment are weighted and aggregated by the attention mechanism, and finally, the olfactory impairment classification prediction result is output through a fully connected layer.

[0119] 1. KAN-GRU module.

[0120] The KAN-GRU module is a deep cyclic temporal encoder composed of multiple KAN-GRU units stacked along the time dimension. Its core innovation lies in using the Kolmogorov-Arnold Network (KAN) to replace the linear transformation in the traditional GRU, thereby computing candidate hidden states in a more flexible nonlinear way.

[0121] Specifically, each KAN-GRU unit retains the reset and update gates of the traditional GRU to maintain the gating mechanism's effective ability to regulate information flow. The computation of candidate hidden states abandons the linear transformation + tanh activation method of the traditional GRU, instead using a KAN layer. The mathematical foundation of the KAN layer comes from the Kolmogorov-Arnold representation theorem, which states that any multivariate continuous function can be represented as a nested superposition of a finite number of univariate continuous functions. Based on this theory, the KAN layer independently maps each dimension of the input vector through a set of learnable univariate B-spline basis functions, enabling a more flexible fit to the complex nonlinear relationship between input and output. This makes it more suitable for capturing complex temporal patterns in EEG signals caused by individual differences, neural conduction delays, and nonlinear coupling.

[0122] Specifically, let the current time be, i.e., the time of the first... The input feature vector at time step (i.e., the fused and reconstructed features) is... ( Features from fusion and reconstruction Vectors segmented by time steps (where the channel number is the number of channels), the previous time step, i.e., the th... The hidden state at time t is ( (To hide the state dimension). The improved KAN-GRU cell computation process is as follows: First, the gating signal for resetting the gate is calculated using a linear transformation of the standard GRU plus a sigmoid activation. and update the door's gate control signal :

[0123]

[0124] in, This indicates vector concatenation; The weight matrix is ​​a learnable matrix; For bias; Use the Sigmoid activation function; For the first Moment, that is, the fusion and reconstruction features of the current moment; For the first Time, that is, the hidden state of the previous time.

[0125] To enhance the model's ability to capture nonlinear dynamic patterns of EEG signals, this invention abandons the traditional GRU approach that relies on... The linear transformation method for calculating candidate hidden states is replaced by a KAN layer. Specifically, each KAN-GRU unit will... The moment, that is, the fusion and reconstruction features of the current moment. After being modulated by the reset gate, the first Time, that is, the hidden state of the previous time. By concatenating the vectors, we obtain the current input vector of the KAN layer. :

[0126] Where ⊙ represents element-wise multiplication; The initial hidden state is a matrix of all zeros.

[0127] The hidden state of the first Dimensions Candidate hidden state The results are obtained by combining learnable univariate spline functions from the KAN layer:

[0128] in, For input vectors The One component; For the first The first input dimension Each B-spline basis function adopts cubic B-spline form, and its shape is determined by a predefined node sequence. After initialization, it can be fine-tuned during training. The number of basis functions configured for a single input dimension; These are learnable weight coefficients used to linearly weight the outputs of each basis function.

[0129] By independently configuring a set of learnable spline basis functions for each input dimension, the KAN layer can fit the complex relationship between the input and the hidden state with a high degree of freedom nonlinear mapping, adapting to the needs of fine temporal feature modeling of EEG signals caused by individual differences, neural conduction delays, and nonlinear coupling of channels.

[0130] Final hidden state:

[0131] Multiple KANs GRU units are stacked along the time dimension to construct a deep cyclic temporal encoder. For the first... A brainwave segment ( The length of its feature sequence is determined by the fusion and reconstruction of features. The feature sequences are divided along the time dimension and then input into the stacked KAN at each time step. GRU unit, take the last KAN Hidden state of the GRU cell at the last time step This is the hidden state of the EEG segment.

[0132] 2. Attention aggregation module.

[0133] The attention aggregation module uses an attention mechanism to weighted aggregate the hidden states of all EEG slices in the EEG slice sequence to obtain global features.

[0134]

[0135] in, For global features; The total number of all EEG slices in the EEG slice sequence; It is the first The hidden state of a computer slice; These are learnable attention weight parameters; This is the index used during aggregation.

[0136] 3. Classification model.

[0137] The classification module obtains the classification and recognition results of the EEG signals based on the global features. Specifically, the global features... The input is a fully connected layer, which is activated via Softmax and outputs the probability distribution of the subjects belonging to each category. :

[0138] in, This is a learnable classification weight matrix; It is the bias vector; The number of categories to be predicted can be set in this invention. These correspond to three levels of grading tasks: healthy, mild olfactory dysfunction, and severe olfactory dysfunction.

[0139] In this invention, supervised training of the improved KAN-GRU classification model can be performed using clinically labeled EEG samples. Specifically, supervised training can be performed using the cross-entropy loss function as the target:

[0140] in, One-hot encoding of the real labels corresponding to the EEG slice series used for training; The probability distributions output by the classification model corresponding to the series of EEG slices used for training.

[0141] During the training of the improved KAN-GRU classification model, all parameters of the dual-path fusion model are frozen, and only the learnable parameters of the improved KAN-GRU classification model are updated, including the GRU gating parameters. KAN layer spline weighting coefficients and the adjustable parameters of the basis functions and the attention aggregation parameters Classification layer parameters and .

[0142] The optimizer can employ the AdamW algorithm, coupled with learning rate preheating and cosine annealing learning rate scheduling strategies, to smoothly converge the process and avoid local optima. Training iterations stop when the classification accuracy on the validation set stabilizes, ultimately constructing a complete EEG recognition model capable of automatically and accurately classifying olfactory disorders.

[0143] Figure 6 A schematic diagram of the olfactory disorder EEG signal recognition system of the present invention is shown. Figure 6 As shown, the olfactory disorder EEG signal recognition system of the present invention includes: 1. EEG signal acquisition and processing module.

[0144] The EEG signal acquisition and processing module is used to acquire the raw multi-channel EEG signals to be processed, and to preprocess and slice the EEG signals to form an EEG slice sequence.

[0145] 2. Fusion and reconstruction feature acquisition module.

[0146] The fusion reconstruction feature acquisition module is used to construct a dual-path fusion model and pre-train the dual-path fusion model. The EEG slice sequence is input into the pre-trained dual-path fusion model to extract fusion reconstruction features. The dual-path fusion model includes an associated topology calibration Mamba model, a multi-prototype Riemann codebook model, and a feature fusion module. The associated topology calibration Mamba model is used to perform temporal modeling through selective state space scanning to extract dynamic temporal features rich in olfactory stimulus response patterns from the EEG slices, and to perform phase alignment of the dynamic temporal features using the associated topology calibration mechanism to obtain temporal features. The multi-prototype Riemann codebook model is used to calculate the covariance matrix of the EEG slices to construct positive definite symmetric manifold nodes, and then project them onto the tangent space and match and weightedly fuse them with the Riemann codebook of learnable Riemann prototypes to obtain prototype features representing individual physiological background. The feature fusion module is used to perform weighted fusion of the temporal features and prototype features to form fusion reconstruction features.

[0147] 3. EEG signal classification and recognition module.

[0148] The EEG signal classification and recognition module is used to input the fused reconstructed features into the improved KAN-GRU classification model after supervised classification training to generate the classification and recognition results of the EEG signals. The improved KAN-GRU classification model includes a KAN-GRU module, an attention aggregation module, and a classification module. The KAN-GRU module is a deep recurrent temporal encoder composed of multiple KAN-GRU units stacked along the time dimension, used to extract the hidden state of each EEG slice based on the fused reconstructed features. The attention aggregation module is used to perform weighted aggregation of the hidden states of all EEG slices in the EEG slice sequence through an attention mechanism to obtain global features. The classification module is used to obtain the classification and recognition results of the EEG signals based on the global features.

[0149] Furthermore, this invention also provides a device for recognizing electroencephalogram (EEG) signals in cases of olfactory dysfunction. For example... Figure 7 As shown, the olfactory disorder EEG signal recognition device of the present invention includes: a memory 11 for storing one or more programs; one or more processors 12; when the one or more programs are executed by the one or more processors 12, the one or more processors 12 implement the olfactory disorder EEG signal recognition method of the present invention. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the olfactory disorder EEG signal recognition method of the present invention.

[0150] The computer-readable storage medium includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined in this invention, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0151] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0152] The steps of the methods or algorithms described in conjunction with the embodiments disclosed in this invention can be implemented in hardware, software modules executed by a processor, or a combination of both. The software modules can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention based on the concept of the present invention, without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for recognizing electroencephalogram (EEG) signals in cases of olfactory dysfunction, characterized in that, Includes the following steps: S1: Acquire the raw multi-channel EEG signal to be processed, and preprocess and slice the EEG signal to form an EEG slice sequence; S2: Construct a dual-path fusion model and pre-train it. Input the EEG slice sequence into the pre-trained dual-path fusion model to extract fusion reconstruction features. The dual-path fusion model includes an associated topology calibration Mamba model, a multi-prototype Riemann codebook model, and a feature fusion module. The associated topology calibration Mamba model is used to perform temporal modeling through selective state space scanning to extract dynamic temporal features rich in olfactory stimulus response patterns from the EEG slices, and uses the associated topology calibration mechanism to perform phase alignment of the dynamic temporal features to obtain temporal features. The multi-prototype Riemann codebook model is used to calculate the covariance matrix of the EEG slices to construct positive definite symmetric manifold nodes, and then project them onto the tangent space and match and weightedly fuse them with the Riemann codebook of learnable Riemann prototypes to obtain prototype features representing individual physiological background. The feature fusion module is used to weightedly fuse the temporal features and prototype features to form fusion reconstruction features. S3: The fused reconstructed features are input into the improved KAN-GRU classification model after supervised classification training to generate the classification and recognition results of the EEG signals. The improved KAN-GRU classification model includes a KAN-GRU module, an attention aggregation module, and a classification module. The KAN-GRU module is a deep recurrent temporal encoder composed of multiple KAN-GRU units stacked along the time dimension, used to extract the hidden state of each EEG slice based on the fused reconstructed features. The attention aggregation module is used to perform weighted aggregation of the hidden states of all EEG slices in the EEG slice sequence through an attention mechanism to obtain global features. The classification module is used to obtain the classification and recognition results of the EEG signals based on the global features.

2. The method for recognizing electroencephalogram (EEG) signals of olfactory disorders according to claim 1, characterized in that, The associated topology calibration Mamba model specifically includes: The selective state-space scanning module is used to recursively scan EEG slices along the channel dimension, in order to To update the formula, the hidden state matrix is ​​iteratively updated using the hidden state matrix from the previous time step and the input data of the current EEG slice. After iterating through the entire EEG slice, the overall hidden state matrix is ​​output. ,in, It is the first The hidden state matrix at time step 1. It is a learnable state transition matrix. It is a learnable input projection matrix. The first EEG slice Input data at any time, It is the first The hidden state matrix at time step 1. The initial hidden state matrix is ​​initialized to an all-zero matrix. A relational topology calibration module, used to calibrate the two-dimensional hidden state matrix. Remodeled into a 3D hidden state matrix ; using the reshaped 3D hidden state matrix A feature association graph is constructed based on the feature association graph. The vertex set of the feature association graph corresponds to all EEG channels, and the edge set represents the functional coupling relationship between channels. Based on the feature association graph, the dynamic adjacency matrix between channels is calculated. Based on the dynamic adjacency matrix, the two-dimensional offset for phase calibration is predicted. Based on the two-dimensional offset, bilinear interpolation is used to correct the position and align the phase of the reconstructed three-dimensional hidden state matrix to obtain the aligned hidden state matrix. The aligned hidden state matrix is ​​restored to two dimensions and mapped through the output projection matrix to obtain the temporal features.

3. The method for recognizing electroencephalogram (EEG) signals of olfactory disorders according to claim 1, characterized in that, The features of the multi-prototype Riemann codebook model specifically include: The tangent space projection and codebook matching module is used to calculate the spatial covariance matrix of EEG slices, and construct positive definite symmetric manifold nodes representing the topological coupling characteristics of neural activity based on the spatial covariance matrix. The positive definite symmetric manifold nodes are projected onto the corresponding Euclidean tangent space using matrix logarithm operations to obtain linearized Riemann representation vectors. The Euclidean distance between the Riemann representation vectors and the prototype vectors in the preset learnable Riemann prototype codebook is calculated. The most similar vector is selected based on the Euclidean distance. A learnable Riemannian prototype; the EEG slice pairs are calculated using a similarity function. Normalized membership weights of a learnable Riemannian prototype; The prototype weighted fusion and geometric restoration module is used to apply the normalized membership weights to the selected... We perform weighted summation and exponentiation on the prototype vectors of learnable Riemann prototypes; then we flatten the features and transform the dimensions of the weighted summation and exponentiation results to obtain the prototype features.

4. The method for recognizing electroencephalogram (EEG) signals of olfactory disorders according to claim 1, characterized in that, During pre-training of the dual-path fusion model, the model parameters are jointly optimized based on reconstruction loss and quantization loss, and the loss function during joint optimization is... for: in, For reconstruction loss; For the first Fusion and reconstruction features of individual EEG slices; For the first Original features of an EEG slice; To quantify the loss; This is the balance coefficient; Indicates the first Riemann representation vectors of computer slices; Indicates the first The prototype features corresponding to each computer slice; This indicates that the gradient operation has been stopped. This represents the Frobenius norm.

5. The method for recognizing electroencephalogram (EEG) signals of olfactory disorders according to claim 1, characterized in that, In step S3, the KAN-GRU module extracts the hidden state of each EEG slice based on the fused reconstruction features as follows: S31: Each of the KAN-GRU units uses a linear transformation of the standard GRU plus a Sigmoid activation to calculate the gate control signal for resetting the gate. and update the door's gate control signal : in, This represents vector concatenation. The weight matrix is ​​a learnable matrix. For bias, Use the Sigmoid activation function; For the first The characteristics of time-matter fusion and reconstruction; For the first The hidden state at all times The initial hidden state is initialized as an all-zero matrix; S32: Each of the KAN-GRU units will... Moment-by-moment fusion and reconstruction features After being modulated by the reset gate, the first Hidden state of time The vectors are concatenated to obtain the input vector of the KAN layer. : Where ⊙ represents element-wise multiplication; S33: The KAN layer of each KAN-GRU unit is based on the input vector. The first is obtained by combining learnable univariate spline functions. Hidden state of time ; S34: The hidden state output by the last KAN-GRU unit at the last moment. This is the hidden state of the EEG slice.

6. The method for recognizing electroencephalogram (EEG) signals of olfactory disorders according to claim 5, characterized in that, In step S3, the attention aggregation module performs weighted aggregation of the hidden states of all EEG slices in the EEG slice sequence through an attention mechanism to obtain global features, specifically: in, As a global feature, The total number of EEG slices in the EEG slice sequence. It is the first The hidden state of a computer slice. These are learnable attention weight parameters. This is the index used during aggregation.

7. The method for recognizing electroencephalogram (EEG) signals of olfactory disorders according to claim 6, characterized in that, In step S3, the classification module obtains the classification and recognition result of the EEG signal based on the global features by: [details of the global features are missing from the original text]. The input is a fully connected layer, which is activated via Softmax and outputs the probability distribution of the subjects belonging to each category. : in, The classification weight matrix is ​​a learnable matrix. This is the bias vector.

8. A brainwave signal recognition system for olfactory disorders, characterized in that, include: The EEG signal acquisition and processing module is used to acquire the raw multi-channel EEG signals to be processed, and to preprocess and slice the EEG signals to form an EEG slice sequence. A fusion reconstruction feature acquisition module is used to construct a dual-path fusion model and pre-train the model. The EEG slice sequence is input into the pre-trained model to extract fusion reconstruction features. The dual-path fusion model includes an associated topology calibration Mamba model, a multi-prototype Riemann codebook model, and a feature fusion module. The associated topology calibration Mamba model performs temporal modeling through selective state space scanning to extract dynamic temporal features rich in olfactory stimulus response patterns from the EEG slices, and uses an associated topology calibration mechanism to phase-align these dynamic temporal features to obtain temporal features. The multi-prototype Riemann codebook model calculates the covariance matrix of the EEG slices to construct positive definite symmetric manifold nodes, projects them onto the tangent space, and then matches and weights them with a Riemann codebook of learnable Riemann prototypes to obtain prototype features representing individual physiological backgrounds. The feature fusion module performs weighted fusion of the temporal features and prototype features to form fusion reconstruction features. An EEG signal classification and recognition module is used to input the fused reconstructed features into an improved KAN-GRU classification model after supervised classification training to generate EEG signal classification and recognition results. The improved KAN-GRU classification model includes a KAN-GRU module, an attention aggregation module, and a classification module. The KAN-GRU module is a deep recurrent temporal encoder composed of multiple KAN-GRU units stacked along the time dimension, used to extract the hidden state of each EEG slice based on the fused reconstructed features. The attention aggregation module is used to perform weighted aggregation of the hidden states of all EEG slices in the EEG slice sequence through an attention mechanism to obtain global features. The classification module is used to obtain the EEG signal classification and recognition results based on the global features.

9. A device for recognizing electroencephalogram (EEG) signals in patients with olfactory disorders, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the electroencephalogram (EEG) signal recognition method for olfactory disorders as described in any one of claims 1-7.

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 steps of the electroencephalogram (EEG) signal recognition method for olfactory disorders as described in any one of claims 1-7.

Images