Adaptive subspace projection suppression method based on external noise of optical pump magnetometer

Abstract

The invention discloses a self-adaptive subspace projection suppression method based on optical pump magnetometer external noise, and belongs to the technical field of neural magnetic imaging (MEG) signal processing.The method comprises the steps that brain magnetic signals are collected through a wearable optical pump magnetometer array and preprocessed, and a signal and noise subspace decomposition model is established; integrating a model-driven noise modeling method and a data-driven noise modeling method, and constructing a noise subspace base vector; constructing a signal subspace basis vector in combination with a guide field matrix and event correlation analysis; identifying residual noise components based on event correlation analysis, and perfecting noise subspace modeling; signal noise separation is realized by adopting a hierarchical projection strategy; and analyzing the time correlation of internal and external space signals, and constructing a time projection matrix to remove residual interference to obtain a clean MEG signal. According to the invention, various types of external noise interferences can be effectively suppressed, the signal-to-noise ratio and the reliability of brain magnetic signals are remarkably improved, and the method is particularly suitable for practical application of a wearable optical pump magnetometer system.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology for neural magnetic imaging (MEG), specifically relating to an adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer. Background Technology

[0002] Magnetoencephalography (MEG) is a high spatiotemporal resolution neuroimaging technique that detects weak magnetic field signals (approximately 100 fT) generated by the activity of neurons in the brain. MEG (Metal-Oriented Genetic Imaging) is used to study brain function at the Tesla level. Compared to electroencephalography (EEG) and functional magnetic resonance imaging (fMRI), MEG has millisecond-level temporal resolution and millimeter-level spatial resolution, making it valuable in areas such as epilepsy localization, cognitive neuroscience research, and brain-computer interfaces. Traditional MEG systems are based on superconducting quantum interference devices (SQUIDs), which require operation in a liquid helium cryogenic environment. This results in drawbacks such as large equipment size, high maintenance costs, and a significant distance (2-3 cm) between the sensors and the scalp, limiting their clinical application.

[0003] In recent years, breakthroughs in optically pumped magnetometer (OPM) technology have driven the development of wearable MEG systems. OPM-MEGs can operate at room temperature, the sensor can be placed directly close to the scalp, and the signal strength is five times higher than that of the SQUID system. They also offer advantages such as low cost, high portability, and flexible deployment. However, OPM-MEG systems are susceptible to interference from environmental magnetic fields (such as the geomagnetic field, power line noise, and mechanical vibration), the intensity of which can reach [amount missing] of the brain's magnetic field signal. This significantly reduces the signal-to-noise ratio.

[0004] Existing noise suppression methods mainly fall into two categories: hardware and software. Hardware methods, such as magnetically shielded rooms and active compensation systems, are effective but costly and limit their application scenarios. Software methods include reference regression, blind source separation, and subspace projection. Among these, subspace projection methods (such as Signal Spatial Projection (SSP) and Uniform Field Corrected (HFC)) achieve interference suppression by separating the signal and noise subspaces. However, traditional methods are based on SQUID system design, rely on high-density sensor arrays (100 channels or more) and strict geometric constraints, making them difficult to adapt to the low channel count (30-60 channels) and variable array layout of OPM-MEG. Improved methods, such as Spatiotemporally Extended HFC (teHFC), are partially compatible with OPM systems, but still suffer from problems such as fixed noise basis, excessive signal loss, or residual noise.

[0005] Therefore, there is an urgent need for a highly adaptive noise suppression method that is compatible with low channel counts, which can combine the advantages of model-driven and data-driven approaches to efficiently separate neural signals from interference in complex environments and promote the clinical adoption of OPM-MEG. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides an adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer. This method can effectively suppress various noise interferences in MEG signals and improve the quality and reliability of MEG signals.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] An adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer, the method comprising:

[0009] Step 1: Collect and preprocess brain magnetic field signals using a wearable optically pumped magnetometer array, establish a signal and noise subspace decomposition model for the MEG observation space, and calculate the projection matrix used for signal separation.

[0010] Step 2: By integrating the model-driven noise modeling method based on sensor geometric information and the data-driven noise modeling method based on empty room information, a preliminary noise subspace basis vector is constructed.

[0011] Step 3: By integrating the model-driven signal modeling method based on the guiding field matrix and the data-driven signal modeling method based on event correlation analysis, construct the signal subspace basis matrix;

[0012] Step 4: Based on the results of event correlation analysis, obtain constant noise patterns that are not induced by stimulating events through a reverse selection strategy, and construct the noise subspace basis matrix;

[0013] Step 5: Based on the constructed signal subspace basis matrix and noise subspace basis matrix, a hierarchical projection strategy is used to achieve signal-noise separation, resulting in internal spatial signals and external spatial signals;

[0014] Step 6: Analyze the temporal correlation of the obtained internal and external space signals, construct a time projection matrix to remove residual interference components, and obtain the final clean MEG signal.

[0015] Furthermore, the process of acquiring brain magnetic field signals using a wearable optically pumped magnetometer array and performing basic preprocessing to establish a signal-noise subspace decomposition model for the MEG observation space and calculate the projection matrix for signal separation includes: acquiring brain magnetic signal data using an optically pumped magnetometer, setting appropriate sampling frequency and recording duration, performing preprocessing operations such as bandpass filtering, notch filtering, and baseline correction on the acquired raw data, checking data quality and marking or removing abnormal channels to obtain the observation data matrix. This includes the number of effective sensor channels. and number of time sampling points Information. The observation data matrix is ​​decomposed into a signal subspace and a noise subspace. The basis matrices of the signal subspace and the noise subspace are obtained through spatial basis representation methods, and a linear representation model of the observed signal is established. ,in and Let be the coefficient vectors of the signal subspace and the noise subspace. and Let be the basis matrix of the subspace of signal and noise, denoted as the parameter to be solved; construct the signal projection matrix. and noise projection matrix It is used to extract signal components from the observation data matrix. and noise components .

[0016] Furthermore, the preliminary noise subspace basis vector is constructed by integrating model-driven noise modeling based on sensor geometry information and data-driven noise modeling based on empty room information, including: obtaining the geometric configuration parameters of the MEG sensor array and constructing the sensor sensing axis direction matrix. For this matrix Eigenvalue decomposition is performed to extract the top-ranked principal eigenvectors, which form the uniform field noise basis. Noise data of the empty room was collected before the formal MEG signal acquisition. The covariance matrix of the noise data from empty rooms is calculated and eigenvalue decomposition is performed. Several top-ranked eigenvectors are selected to form a non-uniform noise basis. The noise basis obtained by model-driven and data-driven methods are fused to construct a preliminary noise subspace basis vector. .

[0017] Furthermore, the construction of signal subspace basis vectors by integrating model-driven signal modeling based on the guiding field matrix and data-driven signal modeling based on event correlation analysis includes: establishing the guiding field matrix by combining sensor location information and magnetic field propagation model. Describes the gain relationship of each spatial voxel to each sensor, and the Gram matrix of the guiding field matrix. Eigenvalue decomposition is performed, and the top-ranked principal eigenvectors are selected to form the basis matrix of the guiding field signal. Using preliminary noise subspace basis vectors Measurement data observation data matrix Perform orthogonal projection to obtain the feature-enhanced data matrix. Construct an event-related time matrix based on experimental stimulus timestamp information. A specific Toeplitz structure is used to encode the temporal pattern; the feature-enhanced data matrix is ​​then processed. Event-related time matrix Perform QR decomposition and construct a first event correlation spatial filter by selecting significant components through singular value decomposition. The signal basis obtained by model-driven and data-driven methods are fused to construct the signal subspace basis matrix. .

[0018] Furthermore, based on the results of the event correlation analysis, residual noise components are further identified and extracted. A reverse selection strategy is used to obtain constant noise patterns uninduced by stimulating events, thus improving the noise subspace modeling. This includes: data... and event matrix Perform QR decomposition and select the right singular vector corresponding to the minimum singular value to construct a second event-correlation spatial filter. The algorithm identifies constant noise components unrelated to event stimuli; it then fuses the initial noise subspace basis vectors with the newly extracted event-independent noise basis to construct a complete noise subspace basis matrix. ,in, Indicates the spatial filter related to the second event. Moore-Penrose pseudo-reverse.

[0019] Furthermore, based on the constructed signal subspace basis and noise subspace basis, a hierarchical projection strategy is employed to achieve refined signal-noise separation, obtaining the internal spatial signal and the external spatial signal, including: utilizing a uniform field spatial basis. For the observation data matrix Orthogonal projection is used to obtain intermediate data Prioritize removing uniform field interference components with clear spatial distribution and significant differences from brain signal activation patterns; construct a signal projection matrix using the constructed signal subspace basis and noise subspace basis. and noise projection matrix This enables fine separation of signal and noise components, yielding the internal spatial signal. and external space signals .

[0020] Furthermore, the temporal correlation of the internal and external space signals obtained from the analysis is used to construct a time projection matrix to remove residual interference components, resulting in the final processed clean MEG signal. This includes: processing the internal space signal separately... and external space signals Perform singular value decomposition, extract the temporal domain feature structures of each, and select the temporal basis vectors corresponding to significantly large singular values ​​to construct the temporal subspace basis matrices, denoted as follows: and ; Correlation matrix of inner and outer time bases Perform singular value decomposition, set a correlation threshold, select highly correlated time components to construct an interference projection operator, and select the time singular vectors corresponding to the singular values ​​that meet the conditions; based on the internal spatial signal... A time-domain projection operator is constructed based on the time base and related components. This projection operator is then used to perform orthogonal projection in the time domain to remove internal spatial signals. The temporal interference component in the signal is removed to obtain the final clean MEG neural signal. .

[0021] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer.

[0022] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer.

[0023] The beneficial effects of this invention are as follows:

[0024] (1) This invention addresses the low channel number and variable array geometry of OPM-MEG systems, effectively suppressing environmental magnetic field interference when the number of channels is limited, significantly improving the signal quality of OPM-MEG systems in non-ideal magnetic shielding environments, and expanding the application scenarios of OPM-MEG.

[0025] (2) This invention establishes a unified noise processing framework that can simultaneously handle environmental magnetic field interference with different characteristics, avoiding the complexity of traditional methods that require separate processing for different noise types, and improving the universality and practicality of the algorithm.

[0026] (3) This invention innovatively combines model-driven basis functions with data-driven adaptive basis functions, which not only utilizes the accuracy of the prior physical model, but also leverages the adaptive capability of the data-driven method, thus preserving the integrity of the neural signal to the greatest extent while ensuring the noise suppression effect.

[0027] (4) The present invention has good algorithm scalability and can easily integrate new prior information and information acquisition methods as technology develops;

[0028] (5) By selectively relaxing the strict orthogonality constraint of the signal / noise subspace, this invention improves the adaptability to complex noise environments while ensuring the stability of the algorithm. Experimental verification shows that compared with the traditional subspace projection method, this invention has significant improvements in both signal-to-noise ratio and neural signal fidelity. Attached Figure Description

[0029] Figure 1 This is a flowchart of the adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to the present invention. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0031] like Figure 1 As shown, the specific implementation process of the adaptive subspace projection method based on external noise of an optically pumped magnetometer proposed in this invention is as follows:

[0032] Step S1: Data acquisition and theoretical modeling;

[0033] Brain magnetic field signals were collected and preprocessed using a wearable optically pumped magnetometer array. A signal and noise subspace decomposition model of the MEG observation space was established, and the projection matrix for signal separation was calculated, providing a data foundation and theoretical tools for subsequent adaptive subspace separation.

[0034] ① Data acquisition and basic preprocessing;

[0035] Magnetoencephalography (MEG) signal data were acquired using an optically pumped magnetometer at a sampling frequency of 1000 Hz. The recording duration was determined based on experimental requirements, typically ranging from 5 to 30 minutes. The raw data underwent the following preprocessing: a 0.1-45 Hz bandpass filter was applied to remove interference outside the physiological signal frequency band; a 50 Hz notch filter was used to eliminate power frequency interference; baseline correction was performed to remove DC offset; and data quality was checked, with abnormal channels marked or removed. The preprocessed data was then recorded as an observation data matrix. ,in For the effective number of sensor channels, This represents the number of time sampling points.

[0036] ② Establishment of subspace decomposition model;

[0037] Establish a subspace decomposition model for the MEG observation space, and... The dimensional observation space is decomposed into two parts: a signal subspace and a noise subspace. Specifically, the basis matrix of the signal subspace (generally denoted as ) is obtained through singular value decomposition or other spatial basis representation methods. (obtained from multiple perspectives in subsequent steps) and the noise subspace basis matrix (generally denoted as...). (obtained from multiple angles in subsequent steps), among which Let be the dimension of the signal subspace. Let be the dimension of the noise subspace, satisfying Ignoring modeling errors, a linear representation model of the observed signal can be established: ,in and These are the coefficient vectors for the signal subspace and the noise subspace, respectively.

[0038] ③ Calculation of the projection matrix;

[0039] Constructing the signal projection matrix and noise projection matrix Used to extract signal components from observed signals. and noise components .

[0040] The specific calculation of the projection matrix can be derived as follows:

[0041] ,

[0042] In the general case of nonorthogonal subspaces, the noise projection operator and the signal projection operator are not complementary, i.e. .like and If the equation is an orthonormal basis, or if the signal subspace and the noise subspace are orthogonal, then the above equation can be simplified.

[0043] Step S2: Preliminary modeling of the noise space;

[0044] By integrating model-driven and data-driven noise modeling methods, a preliminary noise subspace basis vector is constructed, providing a basic noise space representation for subsequent adaptive noise suppression.

[0045] ① Model-driven modeling based on sensor geometric information;

[0046] First, obtain the geometric configuration parameters of the MEG sensor array, which allows you to construct the sensor sensing axis orientation matrix. ,in Indicates that each sensor is in Sensitive axis components in the axial direction, and They represent shaft and Sensitive axial component in the axial direction.

[0047] Based on the far-field assumption, a spatial distribution model of uniform-field interference can be established using the sensitive axis direction matrix. Because when the interference source is sufficiently far from the sensor array, the interference intensity received by each sensor is similar, and the spatial distribution of the interference is mainly determined by the sensor's sensitive axis direction. The extraction method is as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Perform eigenvalue decomposition, where express Moore-Penrose pseudo-reverse:

[0048] ,

[0049] Among them, eigenvalues Arranged in descending order, extract the eigenvectors corresponding to the first three principal eigenvalues ​​to form a uniform field noise basis. .

[0050] ② Data-driven modeling based on vacant room information;

[0051] Based on the additive nature of noise and the assumption of statistical consistency, it is generally assumed that environmental noise during the experiment and noise data collected in an empty room (data collected without subjects) follow the same statistical distribution. Therefore, before the formal MEG signal acquisition experiment begins, it is required to collect noise data in an empty room under the same experimental environment as prior information. Regarding the noise data in the empty room... Calculate the covariance matrix and perform eigenvalue decomposition:

[0052] ,

[0053] in The eigenvalues ​​are sorted in descending order. The top K principal eigenvectors, whose eigenvalues ​​are significantly greater than the others, are selected to form the non-uniform noise basis. .

[0054] ③ Integrate modeling results;

[0055] The noise basis obtained by model-driven and data-driven methods are fused to construct a preliminary noise basis matrix. .

[0056] Step S3: Signal space modeling;

[0057] By integrating model-driven and data-driven signal modeling methods, signal subspace basis vectors are constructed, providing an accurate signal-noise spatial separation foundation for subsequent adaptive noise suppression.

[0058] ① Model-driven signal modeling based on the guiding field matrix;

[0059] First, the source space is modeled, and a guiding field matrix is ​​established by combining sensor location information and magnetic field propagation model. ,in The matrix represents the number of voxel sources. This matrix describes the gain relationship of each spatial voxel to each sensor, characterizing the spatial propagation characteristics of brain source signals. The Gram matrix of the guiding field matrix is ​​decomposed into eigenvalues. ,in The eigenvector matrix, This is an eigenvalue diagonal matrix, with the eigenvalues ​​arranged in descending order. Based on the significant differences in eigenvalues, the dominant signal propagation mode is identified. When there are... When, select the previous The eigenvectors corresponding to the main eigenvalues ​​constitute the basis matrix of the guiding field signal. .

[0060] ② Data-driven signal modeling based on event correlation analysis;

[0061] Considering the insufficient characterization capability of the guiding field matrix when the number of sensors is limited, event-related information is further introduced to optimize the signal subspace construction. To extract the spatial activation mode of the desired signal, the initial noise basis obtained in step S2 is first used. Orthogonal projection of the measurement data is used to achieve feature enhancement:

[0062] ,

[0063] in, It is the identity matrix. This is a data matrix after initial noise suppression.

[0064] Construct an event-related time structure matrix based on experimental stimulus timestamp information. ,in The event response time window length is determined based on the specific experimental paradigm (e.g., a 100-400 ms time window for sensory evoked potentials). This matrix uses a Toeplitz structure to encode event-related time patterns.

[0065] For the data matrix after preliminary noise suppression and structure matrix Perform QR decomposition:

[0066] ,

[0067] in, and It is an orthogonal matrix. and It is an upper triangular matrix.

[0068] Product of orthogonal matrices Perform singular value decomposition ,in and It is an orthogonal matrix. It is a singular value diagonal matrix.

[0069] Choose the right singular vector corresponding to the maximum singular value, and combine it with... Constructing a first event-correlated spatial filter ,in The number of key event-related components selected. express The front of the matrix List.

[0070] The above mathematical transformation process is essentially equivalent to solving the following optimization problem:

[0071] ,

[0072] in, The data components represent the time periods related to the event. The optimization objective is to find a spatial filter that gives the event segment a higher signal-to-noise ratio compared to the non-event segment.

[0073] ③ Integrate modeling results;

[0074] The signal basis obtained by model-driven and data-driven methods are fused to construct the final signal basis matrix. ,in Represents the spatial filter related to the first event. Moore-Penrose pseudo-reverse.

[0075] Step S4: Noise subspace complete;

[0076] Based on the results of the event correlation analysis in step S3, residual noise components are further identified and extracted. A reverse selection strategy is used to obtain constant noise patterns that are not induced by stimulating events, thereby improving the characterization capability of the noise subspace and providing a more accurate noise basis for subsequent adaptive filtering.

[0077] ① Noise component extraction based on event-reverse selection;

[0078] Following the mathematical framework of event correlation analysis in step S3, a reverse selection strategy is used to identify constant noise components unrelated to the event stimuli. Conversely to signal component extraction, the right singular vector corresponding to the minimum singular value is selected to construct a noise component spatial filter (the second event correlation spatial filter). The components corresponding to these minimum singular values ​​represent the spatial patterns least correlated with the event temporal structure, namely, constant background noise unrelated to the experimental stimulus temporal structure, non-event-related components in physiological noise, etc.

[0079] ,

[0080] in The number of main noise components selected, express The matrix corresponds to the minimum A right singular vector with singular values.

[0081] ② Synthetic construction of noise subspace basis;

[0082] The preliminary noise base obtained in step S2 The noise subspace basis matrix is ​​fused with the newly extracted event-independent noise basis from step S4 to construct a complete noise subspace basis matrix:

[0083] ,

[0084] in, Represents the spatial filter matrix related to the second event. Moore-Penrose pseudo-reverse.

[0085] Step S5: Projection of hierarchical space;

[0086] Based on the signal and noise subspace bases constructed in steps S2-S4, a hierarchical projection strategy is employed to achieve refined signal-noise separation. This strategy comprehensively considers the differences in characteristics among different types of noise bases, selectively choosing orthogonal and oblique projection methods to maximize the preservation of target signal components while ensuring noise suppression effectiveness. The orthogonal projection method assumes strict orthogonality between the signal and noise subspaces, achieving noise removal by calculating the orthogonal complement of the noise subspace, thus providing a direct and thorough effect, but carries the risk of inadvertently deleting useful signal components. The oblique projection method, on the other hand, does not require strict orthogonality between the subspaces, achieving a more refined projection by simultaneously considering the geometric relationships between the signal and noise subspaces. This method preserves signal integrity better, but its denoising thoroughness is not as good as orthogonal projection.

[0087] Considering the significant difference between the activation patterns of magnetoencephalography (MEG) signals and the spatial distribution of uniform-field noise bases, and the fact that normal MEG signals do not exhibit uniform-field type spatial activation patterns, uniform-field noise bases can be directly orthogonally projected to obtain the maximum noise suppression benefit. However, the orthogonality characteristics of other noise bases with signal bases are not clear enough, requiring the use of tilted projection methods to avoid signal loss.

[0088] Using uniform field space basis When performing orthogonal projection, prioritize removing the uniform field interference components with the most distinct spatial distribution:

[0089] ,

[0090] in, It is the identity matrix. Let be the projection matrix of the uniform field noise subspace. These are intermediate data after uniform field suppression.

[0091] Using the signal subspace basis calculated in step S3 and residual noise subspace basis (The remaining noise subspace basis is the noise subspace basis matrix calculated in step S4) (After removing the uniform field spatial basis, construct the corresponding signal projection matrix according to the theoretical formula model in step S1). and noise projection matrix This enables fine separation of signal and noise components, yielding the internal spatial signal. and external space signals :

[0092] ,

[0093] Step S6: Temporal domain denoising;

[0094] To further improve signal purification, this invention adds a time-domain denoising step to the spatial domain denoising. Due to the inherent limitations of spatial subspace estimation, the spatial signal within step S5... The signal may still contain temporal components that are unrelated to neural activity but spatially correlated with external interference. This step analyzes the temporal correlation between internal and external signals, constructs a temporal projection matrix to remove residual interference components, and achieves comprehensive purification of the MEG signal.

[0095] ① Time-domain decomposition of inner and outer layer signals;

[0096] Internal space signals respectively and external space signals Perform singular value decomposition to extract the respective time-domain feature structures. The internal spatial signal is decomposed into the following time-domain components:

[0097] ,

[0098] The time-domain decomposition of external space signals is as follows:

[0099] ,

[0100] in, and These are the spatial domain feature vectors of the inner and outer layer signals, respectively. and The corresponding singular values ​​are sorted in descending order. and These are the time-domain feature vectors of the inner and outer layer signals, respectively.

[0101] ② Construction and correlation analysis of temporal subspace basis;

[0102] Based on the energy contribution of singular values, time basis vectors corresponding to significantly large singular values ​​are selected from the internal and external space signals to construct the time subspace basis matrix:

[0103] ,

[0104] Calculate the canonical correlation between subspaces to identify the intersection of two temporal subspaces. Compare the correlation matrices of the inner and outer time bases. Perform singular value decomposition:

[0105] ,

[0106] Among them, singular values Represents the first time interval between two time subspaces. The cosine value of each main character, with a range of values ​​of 1. When the protagonist When approaching zero, A value close to 1 indicates that the two subspaces are highly correlated in that direction. and These are the left and right singular vector matrices, respectively.

[0107] ③ Identification of temporal interference components and construction of projection matrix;

[0108] The interference projection operator is constructed by selecting highly correlated time components. A correlation threshold is set. Generally, the default is Choose to satisfy The number of time singular vectors corresponding to the singular values ​​is denoted as . : .

[0109] Construct the relevant time base matrix A time-domain projection operator is constructed based on the time base and related components of the outer signal:

[0110] ,

[0111] This projection operator can identify interference components in the inner layer signal that are correlated with the outer layer noise in the time domain, and remove them through orthogonal projection. The temporal interference components are removed to obtain purified neural signals.

[0112] ,

[0113] Through the above time-domain denoising process, the MEG signal was further purified, providing a high-quality data foundation for subsequent neural activity analysis and source localization.

[0114] Example:

[0115] This embodiment verifies the technical effectiveness of the MEG signal noise suppression method of the present invention through simulation experiments, and compares and analyzes it with existing methods.

[0116] Experimental Platform Construction: A simulation experimental platform was constructed based on the MNE-Python toolkit. Using T1-MRI data from a 27-year-old male participant, brain tissue was segmented using FreeSurfer software, and a single-layer head model was constructed using the boundary element method to provide an anatomical basis for forward modeling.

[0117] Sensor array configuration: A rigid helmet design employs a 32-channel sensor array, with sensors evenly distributed across the head surface. Precise relative positions between the sensor array and the head are obtained through optical registration technology, constructing a forward-facing model from the neural source to the sensors.

[0118] Source space model: The source space is constructed based on the whole brain cortex, and a point is randomly selected in key functional areas such as the temporal lobe, frontal lobe, occipital lobe and parietal lobe as the dipole source location.

[0119] Simulation signal parameters: The neural signal was set to a 10Hz sine wave, with an activation duration of 0.5 seconds and a resting duration of 0.5 seconds, for a total of 100 trials. The trial interval was a random value of 0.1-0.3 seconds, and the sampling rate was 1000Hz. After projection of the simulation signal onto the forward model, the amplitude was controlled at 400 fT, which is consistent with the range of human brain evoked signals.

[0120] Interference Configuration: To simulate a discrepancy between prior noise information and the actual environment in real-world conditions, the interference source is divided into known and unknown components. The known component uses actual noise from an empty room, while the unknown component is generated by setting two dipoles at the three-dimensional coordinates (0.25m, 0, 0.25m) and (-0.25m, 0, 0.25m) of the head model, producing a 25Hz sine wave signal lasting for 1 second.

[0121] Comparison Methods: The proposed method is compared with existing mainstream subspace projection methods, covering spatial domain, time domain, frequency domain and extended domain methods, mainly including: SSP, CTSP, S3P, HFC, tSSS, DSSP, teHFC and opHFC.

[0122] Evaluation indicators:

[0123] Assuming the ideal brain magnetic signal is The original measurement signal of the sensor is The signal reconstructed by the algorithm is Their scale is In simulation experiments, ideal brain magnetic signals can be obtained. For reference, the following indicators are used for evaluation:

[0124] Shielding Factor (SF): Calculated as the L2 norm ratio between the original measurement signal and the algorithm-reconstructed signal, reflecting the algorithm's ability to suppress noise.

[0125] ,

[0126] Root Mean Square Error (RMSE): Measures the degree of deviation between the signal reconstructed by the algorithm and the ideal magnetoencephalogram (MEG) signal.

[0127] ,

[0128] Relative Noise Reduction (rNR): Represents the relative change in noise amplitude between the original measured signal and the signal reconstructed by the algorithm, used to evaluate the accuracy of noise cancellation.

[0129] ,

[0130] in, This represents the standard deviation calculation. 0% rNR indicates that no noise has been removed, while 100% indicates that the signal has been perfectly reconstructed.

[0131] Signal-to-noise ratio (SNR): Calculates the power ratio between the ideal magnetoencephalogram (MEG) signal and the algorithm-reconstructed signal, reflecting the quality of the algorithm-reconstructed signal.

[0132] ,

[0133] Table 1 shows a comparison of the experimental results for each method.

[0134] Table 1

[0135]

[0136] The indices calculated from the simulation data show that the method of this invention achieves the best performance in all evaluation indicators, including SNR, rNR, RMSE, and SF, compared to other methods, effectively suppressing external interference. The results demonstrate that this invention exhibits excellent robustness in real-world complex noise environments, significantly improving signal quality and providing technical support for the subsequent clinical application and promotion of MEG devices.

[0137] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer.

[0138] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer.

[0139] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer, characterized in that, Includes the following steps: Step 1: Collect and preprocess brain magnetic field signals using a wearable optically pumped magnetometer array, establish a signal and noise subspace decomposition model for the MEG observation space, and calculate the projection matrix used for signal separation. Step 2: By integrating the model-driven noise modeling method based on sensor geometric information and the data-driven noise modeling method based on empty room information, a preliminary noise subspace basis vector is constructed. Step 3: By integrating the model-driven signal modeling method based on the guiding field matrix and the data-driven signal modeling method based on event correlation analysis, construct the signal subspace basis matrix; Step 4: Based on the results of event correlation analysis, obtain constant noise patterns that are not induced by stimulating events through a reverse selection strategy, and construct the noise subspace basis matrix; Step 5: Based on the constructed signal subspace basis matrix and noise subspace basis matrix, a hierarchical projection strategy is used to achieve signal-noise separation, resulting in internal spatial signals and external spatial signals; Step 6: Analyze the temporal correlation of the obtained internal and external space signals, construct a time projection matrix to remove residual interference components, and obtain the final clean MEG signal.

2. The adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to claim 1, characterized in that, Step one includes: Magnetoencephalography (MEG) signal data was acquired using an optically pumped magnetometer, and basic filtering, baseline correction, and bad channel / segment marking were performed to obtain the observation data matrix. ,in For the effective number of sensor channels, This represents the number of time sampling points; The observation data matrix is ​​decomposed into two parts: a signal subspace and a noise subspace. ,in and Let be the coefficient vectors of the signal subspace and the noise subspace. and Let be the basis matrix of the subspace of signal and noise, and let be the parameter to be solved; Constructing the signal projection matrix and noise projection matrix Signal components were extracted from the observation data matrix. and noise components .

3. The adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to claim 1, characterized in that, Step two includes: Obtain the geometric configuration parameters of the MEG sensor array and construct the sensor sensing axis orientation matrix. For the matrix Eigenvalue decomposition is performed, and the top-ranked eigenvectors are extracted to form a uniform field noise basis. , express Moore-Penrose pseudo-inverse; Collect noise data from empty rooms The covariance matrix of the noise data from empty rooms is calculated and eigenvalue decomposition is performed. Several top-ranked eigenvectors are selected to form a non-uniform noise basis. ; uniform field noise base Non-uniform noise base The data is fused to construct a preliminary noise subspace basis vector. .

4. The adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to claim 1, characterized in that, Step three includes: Establish the guiding field matrix by combining sensor location information and magnetic field propagation model Gram matrix of the guiding field matrix Perform eigenvalue decomposition and select the top-ranked eigenvectors to form the basis matrix of the guiding field signal. ; Using the initial noise subspace basis vectors For the observation data matrix Perform orthogonal projection to obtain the feature-enhanced data matrix. ; Construct an event-related time matrix based on experimental stimulus timestamp information. The Toeplitz structure is used to encode the time pattern; Data matrices for feature enhancement Event-related time matrix Perform QR decomposition and select the right singular vector corresponding to the largest singular value to construct the first event correlation spatial filter. ; Guide field signal basis matrix spatial filter related to the first event The obtained signal bases are fused to construct the signal subspace basis matrix. ,in Representation matrix Moore-Penrose pseudo-reverse.

5. The adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to claim 1, characterized in that, Step four includes: Data matrices for feature enhancement Event-related time matrix Perform QR decomposition and select the right singular vector corresponding to the minimum singular value to construct a second event-correlation spatial filter. Identify constant noise components that are unrelated to event stimuli; The initial noise subspace basis vectors are fused with the newly extracted event-independent noise basis to construct the noise subspace basis matrix. ,in, Indicates the spatial filter related to the second event. Moore-Penrose pseudo-reverse.

6. The adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to claim 1, characterized in that, Step five includes: Using uniform field space basis For the observation data matrix Orthogonal projection is used to obtain intermediate data after uniform field suppression. ; Using the constructed noise subspace basis matrix and residual noise subspace basis Constructing the signal projection matrix and noise projection matrix This enables fine separation of signal and noise components, yielding the internal spatial signal. and external space signals The remaining noise subspace basis The basis matrix of the noise subspace The portion after removing the uniform field spatial basis.

7. The adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to claim 1, characterized in that, Step six includes: Internal space signals respectively and external space signals Perform singular value decomposition, extract the temporal domain feature structures of each, and select the temporal basis vectors corresponding to significantly large singular values ​​to construct the temporal subspace basis matrices, denoted as follows: and ; Correlation matrix between internal and external time bases Perform singular value decomposition, set a correlation threshold, select highly correlated time components to construct interference projection operators, and select the time singular vectors corresponding to the singular values ​​that meet the conditions. A time-domain projection operator is constructed based on the time base and related components of the external space signal. This time-domain projection operator is then used to perform orthogonal projection in the time domain to remove time-domain interference components from the internal space signal, resulting in the final clean MEG signal. .

8. The adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer according to claim 7, characterized in that, Among the selected singular values ​​corresponding to the time singular vectors that meet the conditions, the conditions are set as follows: select the singular values ​​that meet the conditions. The singular values ​​correspond to the time singular vectors, where the singular values ​​are... Represents the first of two time subspaces The cosine value of each main character, with a range of values ​​of 1. , This represents the correlation threshold.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor is configured to execute a computer program stored in a memory, which, when executed, implements the adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the adaptive subspace projection suppression method based on external noise of an optically pumped magnetometer as described in any one of claims 1-8.

Images