Signal artifact removal method and system based on multi-band cross-attention mechanism

Abstract

The present invention belongs to the technical field of signal detection, and specifically relates to a signal artifact removal method and system based on a multi-band cross-attention mechanism. The method comprises the following steps: performing multi-band decomposition on a signal to be denoised to obtain a plurality of decomposed signals #imgabs0#, #imgabs1#, wherein #imgabs2# and #imgabs3# represent complete signals to be denoised; performing denoising processing on the plurality of decomposed signals #imgabs4# to obtain a corresponding first output result #imgabs5#; performing multi-band cross-fusion processing on the plurality of first output results #imgabs6# to obtain a second output result #imgabs7#; obtaining a first denoising loss #imgabs9# corresponding to the first output result #imgabs8# and a second denoising loss #imgabs11# corresponding to the second output result #imgabs10# respectively; and obtaining a denoising weighted loss #imgabs14# by combining the first denoising loss #imgabs12# and the second denoising loss #imgabs13# as a model parameter of an artifact removal model. It can effectively learn the signal characteristics of different frequency bands and take into account both the high-frequency and low-frequency components in the signal.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal detection, and in particular relates to a method and system for removing signal artifacts based on a multi-band cross-attention mechanism. Background Art

[0002] EEG is a weak physiological electrical signal. When collecting EEG, it is easy to introduce physiological artifacts such as electrooculography and electromyography, which have a great impact on its subsequent analysis, classification and detection. Specifically, the amplitude of electrooculography artifacts is large and overlaps with the low-frequency band of EEG; electromyography artifacts with high-frequency characteristics will cause the high-frequency band of EEG to be polluted; affected by the contact impedance between electrodes and skin, the signal is also easy to introduce thermal noise interference. Since the amplitude of EEG is inversely proportional to the frequency, if the high-frequency part is not processed well, the high-frequency signal is easy to be lost, resulting in poor fitting performance of the model for the high-frequency components, while the low-frequency components have a greater impact on the model loss.

[0003] In the related technology, for example, patent CN202310112583.6 discloses a deep learning EEG denoising method with dual-branch fusion, in which two branches are designed to simultaneously reconstruct the clean EEG and artifact signals present in the contaminated signal, and a fusion module is used to adaptively fuse the results of the two branches to further restore the clean EEG, thereby improving the quality of the reconstructed EEG signal and greatly improving the denoising ability of the deep neural network. The principle is to use the difference between clean EEG and artifact signals to process and reconstruct them separately. Therefore, the loss function in this technical solution is based on the EEG reconstruction branch, the artifact reconstruction branch, the feature enhancement module and the dual-branch fusion module waiting for the optimization parameters to be obtained, and does not involve the analysis of the impact of the high and low frequencies of the signal on the model and the frequency segmentation processing.

[0004] For example, patent CN202310171586.7 discloses a method for removing EEG artifacts based on a segmentation-denoising network. The segmentation-denoising network model SDNet based on artifact removal is composed of a segmentation subnetwork and a denoising subnetwork. It can judge whether the EEG segment needs denoising based on the artifact position information extracted by the segmentation subnetwork, and automatically reconstruct high-quality clean EEG signals. Although a one-dimensional semantic segmentation network is designed for artifact evaluation, its principle is to judge whether denoising is needed by distinguishing between noisy EEG segments and clean EEG segments and artifact position information, that is, "a one-dimensional semantic segmentation network for artifact evaluation is designed, which can extract shallow and deep semantic features in EEG signals, and distinguish between noisy EEG segments and clean EEG segments by decoding the above semantic information, so that denoising can be performed on noisy EEG segments, overcoming the information loss caused by denoising clean segments by traditional denoising methods, and improving the quality of EEG signals." Summary of the invention

[0005] The present invention provides a signal artifact removal method and system based on a multi-band cross-attention mechanism, which obtains noise reduction loss by multi-band decomposition as a model parameter, thereby solving the different effects of high-frequency signals and low-frequency signals on the model.

[0006] In order to solve the above technical problems, the present invention provides a method for obtaining model parameters, comprising: multi-band decomposition, obtaining multiple decomposed signals based on the signal to be denoised , ,in , Represents the full-band signal to be denoised; denoising processing, that is, multiple decomposed signals Perform noise reduction processing respectively to obtain the corresponding first output result ; Multi-band cross-fusion processing, that is, based on multiple first output results Get the second output result ; Obtain the noise reduction loss, including: obtaining the first output result respectively The corresponding first noise reduction loss , the second output result The corresponding second noise reduction loss ; Get model parameters and combine the first denoising loss , Second noise reduction loss Get the denoising weighted loss , as the model parameters of the artifact removal model.

[0007] Furthermore, the noise reduction process includes: After passing through the convolutional neural network, the features are obtained , ; The feature After N layers of residual modules and a convolutional neural network processing, the features are obtained. ; The feature With features After adding, the first output result is obtained through a convolution layer. .

[0008] Furthermore, the multi-band cross-fusion processing includes: band splicing, combining the multiple first output results Composition of signal matrix ,Right now ; Cross attention mechanism, the signal matrix Get each decomposition signal through the cross attention module Correlated coupled signals ; Band fusion, coupled signal The frequency band is fused by convolution operation; the signal is reconstructed to obtain the second output result .

[0009] Furthermore, the cross-attention module is configured as a Transformer model, that is, the signal matrix Each element in one sequence is associated with each element in another sequence to obtain the coupling relationship between different decomposed signals.

[0010] Further, the obtaining of the noise reduction loss comprises: respectively calculating the first output result The first noise reduction loss is obtained by comparing the root mean square error of the corresponding clean signal with . Calculate the second output result The second noise reduction loss is obtained by comparing the root mean square error of the clean signal with .

[0011] Furthermore, the noise reduction weighted loss Configured as + ,in Decomposition signal The corresponding denoising loss weight.

[0012] Further, the multi-band decomposition includes: at least one of filtering decomposition, wavelet decomposition, and empirical mode decomposition; wherein the wavelet decomposition includes: setting the minimum frequency span in the wavelet decomposition not to exceed the minimum rhythm of the target signal in the selected signal to be denoised; obtaining the number of wavelet decomposition layers of the signal to be denoised based on the minimum frequency span ,set up -1, Indicates integer operation. Indicates the minimum frequency span of the signal to be denoised; represents the sampling rate of the signal to be denoised; performing wavelet decomposition on the signal to be denoised based on the number of wavelet decomposition layers to obtain the multiple decomposed signals .

[0013] In a second aspect, the present invention provides a method for constructing an artifact removal model, including: model training, training verification through training signals in a database; model parameter optimization, that is, optimizing the parameters of the artifact removal model based on different model parameters; the model parameters include the denoising weighted loss obtained by the acquisition method .

[0014] Furthermore, the model parameters also include at least one of the convolution kernel size, the number of convolution layers, and the number of feature maps in the convolutional neural network model.

[0015] Further, the training signal is configured as signal data after data distribution expansion; the data distribution expansion includes at least one of frequency distribution expansion and re-reference expansion; wherein the frequency distribution expansion includes: subjecting the artifact signal and the target signal in the simulation signal to multiple types of frequency distribution expansion respectively to obtain the signals to be denoised with different frequency distributions; wherein the frequency distribution expansion includes: at least one of low-pass filtering, high-pass filtering, notch filtering, re-reference filtering, and spatial domain filtering; the filter gear step of the frequency distribution expansion is configured as the transition bandwidth of the artifact signal; the re-reference expansion includes: setting the extended data , I, t=1,2,3....,T, N is the number of channels, T is the total length of data, For the specified channel Or the average result of the channel is , is the number of channels participating in the averaging.

[0016] In a third aspect, the present invention provides a method for removing signal artifacts, comprising: using the construction method to construct at least one type of artifact removal model; slicing the acquired signal to obtain multiple signal time windows; model selection, using the artifact detection model to detect the artifact type to which the signal time window belongs to, so as to call the corresponding type of artifact removal model; iteratively removing artifacts.

[0017] Furthermore, the artifact removal model includes at least one of a low-frequency signal artifact removal model, a high-frequency signal artifact removal model, and a Gaussian noise artifact removal model.

[0018] In a fourth aspect, the present invention provides a signal artifact removal system, comprising: a processor, an artifact detection model and at least one type of artifact removal model; wherein the processor runs the acquisition method to set the model parameters of the artifact removal model; or the processor runs the construction method to construct the artifact removal model; or the processor runs the removal method to remove artifacts in the acquired signal.

[0019] In a fifth aspect, the present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0020] In a sixth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the method when executed by a processor.

[0021] In a seventh aspect, the present invention provides a computer program product, comprising a computer program, which implements the steps of the method when executed by a processor.

[0022] The beneficial effect of the present invention is that the method for obtaining the model parameters of the present invention first obtains multiple decomposition signals by multi-band decomposition; then performs noise reduction processing on the multiple decomposition signals to obtain the first noise reduction loss of the multiple decomposition signals, and then performs multi-band cross-fusion processing on the multiple decomposition signals to obtain a second output result to obtain the second noise reduction loss of the full-band signal, and finally combines the first noise reduction loss and the second noise reduction loss to obtain the noise reduction weighted loss as the model parameter, which can effectively learn the signal characteristics of different frequency bands, can take into account the high-frequency components and low-frequency components in the signal, and at the same time consider the coupling relationship between signals of different frequency bands, so that the noise reduction effect is closer to the real EEG signal characteristics.

[0023] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0024] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0026] Figure 1 It is a principle block diagram of the method for obtaining model parameters of the present invention.

[0027] Figure 2 It is the decomposed signal result of wavelet decomposition in the multi-band decomposition of the present invention.

[0028] Figure 3 It is a schematic diagram of the model framework of the noise reduction model of the present invention.

[0029] Figure 4 It is a schematic diagram of the model architecture of the convolutional neural network in the noise reduction model of the present invention.

[0030] Figure 5 It is a schematic diagram of the model architecture of the residual module in the noise reduction model of the present invention.

[0031] Figure 6 It is a principle block diagram of the data distribution expansion method of the present invention.

[0032] Figure 7 is the result of different frequency distributions of noisy signals.

[0033] Figure 8 is the result of different frequency distribution of clean signal.

[0034] Fig. 9 It is a principle block diagram of the artifact removal model of the present invention.

[0035] Fig.10 It is a flow chart of the method for removing signal artifacts of the present invention.

[0036] Fig.11 It is a comparison diagram of the power spectral density of the signal before and after noise reduction without data expansion distribution in experimental condition 1.

[0037] Fig.12 This is a comparison of the signal waveforms before and after noise reduction without data expansion distribution in experimental condition 1.

[0038] Fig.13 It is a comparison diagram of the power spectral density of the signal before and after noise reduction after data expansion distribution in experimental condition 2.

[0039] Fig.14 This is a comparison diagram of the signal waveform before and after noise reduction after data expansion distribution in experimental condition 2.

[0040] Fig.15 It is the clean EEG signal in Experiment 2.

[0041] Fig.16 It is the signal containing electromyographic noise after the electromyographic artifact is added to the clean EEG signal in Experimental Example 2.

[0042] Fig.17 is the signal after removing the electromyographic artifact in experimental condition 3.

[0043] Fig.18 is the signal after removing the electromyographic artifact in experimental condition 4. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Example 1.

[0046] See Figure 1-Figure 5, this embodiment 1 provides a method for obtaining model parameters based on a multi-band cross attention mechanism, including: multi-band decomposition, obtaining multiple decomposed signals based on the signal to be denoised , ,in , Represents the full-band signal to be denoised; denoising processing, that is, multiple decomposed signals Perform noise reduction processing respectively to obtain the corresponding first output result ; Multi-band cross-fusion processing, that is, based on multiple first output results Get the second output result ; Obtain the noise reduction loss, including: obtaining the first output result respectively The corresponding first noise reduction loss , the second output result The corresponding second noise reduction loss ; Get model parameters and combine the first denoising loss , Second noise reduction loss Get the denoising weighted loss , as the model parameters of the artifact removal model.

[0047] Optionally, the denoising weighted loss Configured as + ,in Decomposition signal Corresponding noise reduction loss weight. Specifically, different artifact removal models can be weighted according to the type of noise reduction signal. Low-frequency noise gives a large weight to the low-frequency band of the low-frequency signal artifact removal model, high-frequency noise gives a large weight to the high-frequency band of the high-frequency signal artifact removal model, and Gaussian noise is weighted on average to the Gaussian noise artifact removal model, and the total weight is 1.

[0048] Optionally, the signal to be denoised is, for example but not limited to, a training signal in a database, such as electrophysiological signals such as EEG, EMG, EOG, and EKG. For example, EEG is used as the target signal, and other signals are noise signals.

[0049] As an optional implementation of multi-band decomposition.

[0050] The multi-band decomposition includes at least one of filtering decomposition, wavelet decomposition, and empirical mode decomposition (EMD-like decomposition). Optionally, filtering decomposition and empirical mode decomposition are, for example but not limited to, decomposition methods using conventional techniques, where the signal to be denoised is decomposed in the time domain to obtain a decomposed signal including a full-band signal and multiple frequency bands (such as delta, theta, alpha, beta, gamma, etc.).

[0051] Optional, see Figure 2 The wavelet decomposition includes: the minimum frequency span in the wavelet decomposition does not exceed the minimum rhythm of the target signal in the selected signal to be de-noised; based on the minimum frequency span, the number of wavelet decomposition layers of the signal to be de-noised is obtained. ,set up -1, Indicates integer operation (including rounding up or down), A S Indicates the minimum frequency span of the signal to be denoised; represents the sampling rate of the signal to be denoised; performing wavelet decomposition on the signal to be denoised based on the number of wavelet decomposition layers to obtain the multiple decomposed signals ,but For example, the sampling rate of the signal to be denoised , the minimum frequency span of wavelet stratification , the number of wavelet decomposition layers The stratification results are three layers: the first layer , second floor 25-50 , the third layer 0-12.5 and 12.5-25 ;but , the frequency range of the signal to be denoised is 0- , can be decomposed into 5 (from 0-4) decomposition signals , and its corresponding frequency range is as follows: Decomposed signal 0- , decompose the signal 0-12.5 , decompose the signal 12.5- , decompose the signal For 25- , decompose the signal 50- It can be seen that the frequency range of each decomposed signal is not evenly distributed, and the corresponding bandwidth is not the same. Since the low-frequency component has a greater impact on the model loss, the low-frequency signal is decomposed more finely, which is conducive to improving the accuracy of noise reduction in each decomposed signal, thereby improving the first noise reduction loss. The decomposed signal of the high-frequency part has a larger bandwidth and is more suitable for the model to fit the second noise reduction loss corresponding to the high-frequency component. Taking EEG signal detection artifact as an example, the target signal is the EEG signal, and its rhythm can generally be set to Delta (0-4Hz), Theta (4-8Hz), Alpha (8-12Hz), Beta (12-30Hz), Gamma (>30Hz). However, the amplitude of the electrooculogram artifact is large and the frequency range is narrow, so a relatively wide minimum frequency span A can be set. S, such as 4Hz. However, the wide-band EMG artifacts overlap more with the high-frequency band of the EEG signal, so a relatively narrow minimum frequency span A can be set. S , such as 12Hz.

[0052] As an optional implementation manner of obtaining the first noise reduction loss.

[0053] See Figure 3-Figure 5 , obtain multiple decomposed signals The first noise reduction loss includes: noise reduction processing, decomposing multiple signals De-noising is performed using the denoising model to obtain the corresponding first output result ; Calculate the first output result The root mean square error of the corresponding clean signal is the first noise reduction loss .

[0054] Optionally, the denoising model is configured as a deep network model designed for decomposed signals in different frequency ranges, i.e., a residual module (e.g. Figure 5 The number of layers N and the size of the convolution kernel are set according to the frequency range of the decomposed signal. Figure 3 The specific noise reduction process includes: decomposing the signal After convolutional neural network (CNN Block, such as Figure 4 After that, the feature ; The feature After being processed by N layers of residual modules (i.e., ResBlock layers) and a convolutional neural network, the features are obtained. ; The feature With features After adding, the first output result is obtained through a convolution layer. .

[0055] As an optional implementation manner of obtaining the second noise reduction loss.

[0056] See Figure 1 , get the second output result The corresponding second noise reduction loss Including: sub-band splicing, multiple first output results Splicing into signal matrix ,in , signal matrix The dimension of should meet the dimension requirements of the cross attention module; the cross attention mechanism converts the signal matrix The coupled signals of sub-band correlation are obtained through the cross-attention module ; Band fusion: the coupled signal The frequency band is fused through convolution operation; the signal is reconstructed to obtain the second output result ; Calculate the second output result The second noise reduction loss is obtained by comparing the root mean square error of the corresponding clean signal with .

[0057] Optionally, the cross attention module is configured as a Transformer model, that is, the signal matrix Each element in one sequence is associated with each element in another sequence to obtain the coupling relationship between different decomposed signals. The multi-band cross-attention module can focus on the signals of multiple frequency components, avoid transition smoothness and loss of high-frequency signals, and effectively learn the signal characteristics of different frequency bands. The cross-attention mechanism also takes into account the coupling relationship between signals of different frequency bands. Combined with the different frequency ranges of high-frequency and low-frequency components of each decomposed signal, it avoids only considering low-frequency and large-amplitude signals while ignoring the fitting of small-amplitude high-frequency signals during the denoising model learning process, making the denoising effect closer to the real EEG signal characteristics.

[0058] In this embodiment, the clean signal is a simulated signal, such as but not limited to a simulated EEG signal.

[0059] Example 2.

[0060] Based on Example 1, see Figure 6-Figure 9 This embodiment 2 provides a method for constructing an artifact removal model, including: model training, training verification through training signals in a database; model parameter optimization, that is, optimizing the parameters of the artifact removal model based on different model parameters to obtain the best model parameters; the model parameters include the denoising weighted loss obtained by the acquisition method Optionally, the model parameters also include at least one of the convolution kernel size, the number of convolution layers, and the number of feature maps in the convolutional neural network model. The model parameter space is optimized by a greedy algorithm to obtain the optimal model parameters under different artifact types, and complete the data training of multiple types of artifact removal models.

[0061] Further, the training signal is configured as signal data after data distribution expansion; the training signal is configured as signal data after data distribution expansion; the data distribution expansion includes at least one of frequency distribution expansion and re-reference expansion.

[0062] Optionally, the frequency distribution expansion includes: the artifact signal and the target signal in the simulation signal are respectively subjected to multiple types of frequency distribution expansion to obtain the signals to be denoised with different frequency distributions, and the artifact-clean data under various signal processing conditions are used as training sets to improve the overfitting problem of the deep denoising model on the filter characteristics; wherein the frequency distribution expansion includes: at least one of low-pass filtering, high-pass filtering, notch filtering, re-reference filtering, and spatial filtering; the filter gear step of the frequency distribution expansion is configured as the transition bandwidth of the artifact signal. Specifically, based on the simulated EEG signal model, a large number of training databases containing multiple types of artifact signals and clean signals for training verification are generated; by simultaneously processing the target signal and the artifact signal in the database through low-pass and / or high-pass filtering, the signals with different frequency distributions are increased to form an extended database. The finer the frequency distribution gear, the wider the signal distribution, and the frequency gear step is preferably greater than the transition band width of the artifact signal. Moreover, the gears of different artifact types are also different, and the broadband artifact has more gears and larger steps. Taking noisy EEG signals and clean EEG signals as various types of training signals as examples, their frequency distribution expansion method adopts low-pass filtering under filtering conditions 1-6, and the filtering gears are shown in Table 1. The different frequency distribution results of noisy signals are shown in Table 1. Figure 7 As shown; the different frequency distribution results of the clean signal are as follows Figure 8 As shown in Figure 2, both are frequency distributions after filtering. Compared with the frequency distribution of the signal before filtering, the frequency distribution of the signal after filtering is more.

[0063] Table 1 Low-pass filter position

[0064]

[0065] Optionally, the re-reference extension includes: setting extension data , I, t=1,2,3....,T, N is the number of channels, T is the total length of data, For the specified channel Or the average result of the channel is , is the number of channels participating in the averaging.

[0066] Example 3.

[0067] Based on Example 1 or 2, see Fig. 9 and Fig.10 , this embodiment 3 provides a method for removing signal artifacts, including: using the construction method to construct at least one type of artifact removal model; slicing the collected signal to obtain multiple signal time windows; model selection, using the artifact detection model to detect the artifact type to which the signal time window belongs, so as to call the corresponding type of artifact removal model; iteratively removing artifacts. Specifically, see Fig.10 , slice the actual acquisition signal and divide it into multiple time windows, standardize the data in each time window, input the standardized data into the artifact detection model, determine the artifact type of the time window signal, call the corresponding artifact removal model according to the artifact type of the time window, iterate the process until the artifact removal is completed, destandardize the standard signal after artifact removal, and obtain a clean signal after artifact removal; repeat the above operation for each time window to complete the artifact removal processing of the acquisition data.

[0068] Optional, see Fig. 9 The artifact removal model includes: at least one of a low-frequency signal artifact removal model, a high-frequency signal artifact removal model, and a Gaussian noise artifact removal model.

[0069] Example 4.

[0070] Based on Example 3, this Example 4 provides a signal artifact removal system, including: a processor, an artifact detection model and at least one type of artifact removal model; wherein the processor runs the acquisition method to set the model parameters of the artifact removal model; or the processor runs the construction method to construct the artifact removal model; or the processor runs the removal method to remove artifacts in the acquired signal.

[0071] Example 5.

[0072] On the basis of Embodiment 1, 2 or 3, this Embodiment 5 provides a computer device, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0073] Example 6.

[0074] On the basis of Embodiment 1, 2 or 3, this Embodiment 6 provides a computer-readable storage medium on which a computer program is stored, and the computer program implements the steps of the method when executed by a processor.

[0075] Example 7.

[0076] On the basis of Embodiment 1, 2 or 3, this Embodiment 7 provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the method are implemented.

[0077] Experimental Example 1: Data distribution expansion.

[0078] Test condition 1: see Fig.11 and Fig.12The denoising model is the EEGANet EMG denoising model that is not trained with data distribution expansion data. The test signal selects a real EEG signal containing EMG artifacts. The evaluation indicators are the degree of attenuation of the time domain waveform noise of the signal before and after denoising and the degree of low-frequency retention and high-frequency attenuation of the power spectrum of the signal before and after denoising.

[0079] Test condition 2: see Fig.13 and Fig.14 The denoising model is the EEGANet EMG denoising model trained with data distribution extension data. The test signal selects a real EEG signal containing EMG artifacts. The evaluation index is to observe the attenuation degree of the time domain waveform noise of the signal before and after denoising and the low-frequency retention and high-frequency attenuation degree of the power spectrum of the signal before and after denoising.

[0080] Conclusion: By expanding the data distribution, the model learns more robust EEG distribution features rather than just filter features.

[0081] Experimental Example 2: Electromyographic artifacts.

[0082] In the clean EEG signal (see Fig.15 ) is added with myoelectric artifacts to obtain a signal containing myoelectric noise (see Fig.16 ), as the noisy signal before noise reduction under experimental conditions 3 and 4, to compare the effects of the present invention and conventional technology on removing electromyographic artifacts.

[0083] Test condition 3: see Fig.17 The EEGAnet model is selected as the model, the convolution kernel size is set to 1*9, the number of feature maps is set to 64, the number of Resblock layers is set to 16, and the training data is selected from the simulated EEG database containing EMG artifacts and clean data pairs. The sampling rate is 1000Hz, and the test data is also simulated EEG data mixed with EMG artifacts. The evaluation indicators are the attenuation degree of artifacts in the signal before and after denoising, the correlation coefficient between the denoised signal and the clean signal, and the relative root mean square error value.

[0084] Test condition 4: see Fig.18 The model selects the multi-band cross attention model, the convolution kernel size is set to 1*9, the number of feature maps is set to 64, the number of residual modules is set to 4, the minimum frequency span is set to 12Hz, and the loss weight is set to be greater than the high-frequency band weight. The training data selects the pairs of data containing electromyographic artifacts and clean data in the simulated EEG database, with a sampling rate of 1000Hz. The test data is also simulated EEG data with mixed electromyographic artifacts. The evaluation indicators are the attenuation degree of artifacts in the signal before and after denoising, the correlation coefficient between the denoised signal and the clean signal, and the relative root mean square error value.

[0085] Conclusion: After removing the electromyographic artifacts, the signal of the present invention is significantly better than the clean signal in terms of consistency. The higher the CC (correlation coefficient) after denoising and the smaller the RRMS (relative root mean square error), the better the performance of the denoising algorithm. The specific CC and RRMS after denoising are shown in Table 2.

[0086] Table 2 Comparison of noise reduction indicators for electromyographic artifacts

[0087]

[0088] In the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0089] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0090] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0091] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for obtaining model parameters, characterized in that: include: Multi-band decomposition, obtaining multiple decomposed signals based on the signal to be denoised , ,in , Represents the signal to be de-noised in the full frequency band; Noise reduction processing, that is, multiple decomposed signals Perform noise reduction processing respectively to obtain the corresponding first output result ; Multi-band cross-fusion processing, that is, based on multiple first output results Get the second output result ; Obtain the noise reduction loss, including: obtaining the first output result respectively The corresponding first noise reduction loss , the second output result The corresponding second noise reduction loss ; Get model parameters and combine the first denoising loss , Second noise reduction loss Get the denoising weighted loss , as the model parameters of the artifact removal model; The noise reduction process comprises: Multiple decomposed signals After passing through the convolutional neural network, the features are obtained ; The characteristics After N layers of residual modules and a convolutional neural network processing, the features are obtained. ; The characteristics With features After adding, the first output result is obtained through a convolution layer. ; The multi-band cross-fusion process includes: Band splicing, multiple first output results Composition of signal matrix ,Right now ; Cross-attention mechanism, the signal matrix Get each decomposition signal through the cross attention module Correlated coupled signals ; Band fusion, coupled signal Frequency band fusion is accomplished through convolution operations; Signal reconstruction, the second output result is obtained ; The obtaining of the noise reduction loss comprises: Calculate the first output result respectively Decomposed signal with corresponding clean signal The root mean square error is obtained to obtain the first noise reduction loss ; Calculate the second output result The second noise reduction loss is obtained by comparing the root mean square error of the corresponding clean signal with .

2. The acquisition method according to claim 1, characterized in that: The cross-attention module is configured as a Transformer model, that is, the signal matrix Each element in one sequence is associated with each element in another sequence to obtain the coupling relationship between different decomposed signals.

3. The acquisition method according to any one of claims 1 to 2, characterized in that: The denoising weighted loss Configured as + ,in Decomposition signal The corresponding denoising loss weight.

4. The acquisition method according to claim 1, characterized in that: The multi-band decomposition includes at least one of filtering decomposition, wavelet decomposition, and empirical mode decomposition; wherein The wavelet decomposition comprises: The minimum frequency span in the wavelet decomposition is set not to exceed the minimum rhythm of the target signal in the selected signal to be denoised; Get the number of wavelet decomposition layers of the signal to be denoised based on the minimum frequency span ,set up -1, Indicates integer operation. Indicates the minimum frequency span of the signal to be denoised; Indicates the sampling rate of the signal to be denoised; Based on the number of wavelet decomposition layers, the denoised signal is subjected to wavelet decomposition to obtain the multiple decomposed signals. .

5. A method for constructing an artifact removal model, characterized in that: include: Model training, training verification through training signals in the database; Model parameter optimization, that is, optimizing the parameters of the artifact removal model based on different model parameters; The model parameters include the denoising weighted loss obtained by the acquisition method as claimed in claim 1 .

6. The construction method according to claim 5, characterized in that: The model parameters also include at least one of the convolution kernel size, the number of convolution layers, and the number of feature maps in the convolutional neural network model.

7. The construction method according to claim 5, characterized in that: The training signal is configured as signal data after data distribution expansion; The data distribution extension includes at least one of a frequency distribution extension and a re-reference extension; in The frequency distribution expansion includes: subjecting the artifact signal and the target signal in the simulation signal to multiple types of frequency distribution expansion respectively, so as to obtain the signals to be denoised with different frequency distributions; wherein the frequency distribution expansion includes: at least one of low-pass filtering, high-pass filtering, notch filtering, re-reference filtering, and spatial domain filtering; the filter gear step of the frequency distribution expansion is configured as the transition bandwidth of the artifact signal; The re-reference extension includes: setting extension data , I, t=1,2,3....,T, N is the number of channels, T is the total length of data, For the specified channel Or the average result of the channel is , is the number of channels participating in the averaging.

8. A method for removing signal artifacts, characterized in that: include: Constructing at least one type of artifact removal model using the construction method as claimed in claim 5; Slice the collected signal to obtain multiple signal time windows; Model selection: using the artifact detection model to detect the artifact type to which the signal time window belongs, so as to call the corresponding type of artifact removal model; Iterative artifact removal.

9. The removal method according to claim 8, characterized in that: The artifact removal model includes at least one of a low-frequency signal artifact removal model, a high-frequency signal artifact removal model, and a Gaussian noise artifact removal model.

10. A signal artifact removal system, characterized in that: include: a processor, an artifact detection model, and at least one type of artifact removal model; in The processor runs the acquisition method according to claim 1 to set the model parameters of the artifact removal model; or The processor executes the construction method as claimed in claim 5 to construct the artifact removal model; or The processor executes the removal method according to claim 8 to remove artifacts from the acquired signal.

11. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to claim 1 or 5 or 8.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to claim 1 or 5 or 8 are implemented.

13. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to claim 1 or 5 or 8 are implemented.

Images