Low-frequency noise signal detection method

Abstract

The present invention discloses a low-frequency noise signal detection method, comprising: a signal time window division model for dividing the signal time window; a signal time window detection model for dividing the signal time window into a noise signal time window and a normal signal time window; and a signal processing model for separating the low-frequency noise signal from the noise signal time window. The present invention employs a combined time-space domain approach to measure, classify, locate, and separate multiple types of noise signals in a signal. This method can accurately measure and separate various types of noise signals, reduce the loss of effective signals during the noise reduction process, and improve the signal-to-noise ratio.

Description

[0001] This application is a divisional application of the Chinese patent application with the application date of June 6, 2023, application number 202310658542.7, and invention name: Signal noise detection method, signal detection model, and readable storage medium. Technical Field

[0002] The present invention relates to the technical field of signal detection, and in particular to a low-frequency noise signal detection method. Background Art

[0003] Because the acquisition process of the original signal is often accompanied by other noise signals, resulting in a low signal-to-noise ratio and significantly affecting signal quality, noise reduction is therefore a necessary preprocessing step before application. The noise reduction process generally involves locating and removing noise signals. Methods such as ICA (Independent Component Analysis) and PCA (Principal Component Analysis) are typically used to locate and analyze the noise signal, followed by filtering to remove it. However, these methods have two major shortcomings: First, they process the decomposed components too roughly, often directly discarding entire sub-components containing noise signals, resulting in the elimination of some useful signals; second, they target a single type of noise signal; even when the signal contains multiple types of noise, the same processing method is still used. Summary of the Invention

[0004] The present invention aims to solve one of the technical problems existing in the prior art. To this end, the present invention provides a low-frequency noise signal detection method that can improve the signal-to-noise ratio of a reference signal by measuring, classifying, evaluating, locating, and separating the noise signal from the original signal.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a low-frequency noise signal detection method, comprising:

[0006] Signal time window division model, divides the signal time window;

[0007] Signal time window detection model, which divides the signal time window into noise signal time window and normal signal time window;

[0008] A signal processing model is used to separate the low-frequency noise signal in the noise signal time window.

[0009] Furthermore, the signal processing model includes:

[0010] Acquire a spatial decomposition matrix U and a temporal decomposition matrix V based on the noise signal time window;

[0011] Extract the first M columns from the time decomposition matrix V to form the source component matrix A, where M is the number of channels;

[0012] Locate the position of the low-frequency noise signal in the source component matrix A;

[0013] Determine whether the position of the low-frequency noise signal in the source component matrix A is a null value;

[0014] When the position of the low-frequency noise signal is a non-null value, separating the low-frequency noise signal and reconstructing the source are sequentially performed to continue iteration;

[0015] When the position of the low-frequency noise signal is a null value, the iteration is terminated.

[0016] Furthermore, the spatial decomposition matrix U is an M×M matrix, and the temporal decomposition matrix V is a K×K matrix, where K is the number of time domain sampling points.

[0017] Furthermore, the source reconstruction includes:

[0018] Output the matrix V_hat after filtering the low-frequency noise signal;

[0019] An inverse source decomposition operation is performed on the spatial decomposition matrix U and the matrix V_hat to obtain a signal S_hat.

[0020] Furthermore, the position of the low-frequency noise signal in the positioning source component matrix A includes:

[0021] Performing standardization on the source component matrix A to obtain a source component matrix B;

[0022] Performing peak statistics on each column of signals in the source component matrix B to obtain the number of peak values of each column of signals;

[0023] Calculate the peak frequency p of each column signal per unit time m1 ;

[0024] If the peak frequency p m1 > the set frequency threshold p0, the corresponding column signal is marked as a low-frequency noise signal sub-component, and the time position L of the low-frequency noise signal sub-component is recorded. v ;

[0025] After mapping, each column signal of the spatial decomposition matrix U is matched with the residual spatial position template to obtain the matching coefficient;

[0026] If the matching coefficient is greater than the set matching threshold, the spatial position L of the low-frequency noise signal subcomponent is output. u ;

[0027] According to the formula L l =merge(L v ,L u ) Determine the location of the low-frequency noise signal L l .

[0028] Furthermore, separating the low-frequency noise signal includes:

[0029] For position L l The low-frequency interference sub-components on the image are processed by wavelet threshold filtering to separate the low-frequency noise signal.

[0030] Furthermore, the normalization process of the source component matrix A includes:

[0031] Calculate the mean μ of each column signal in the source component matrix A m =mean(v m ) and variance σ m =std(v m ), m=1, 2, ..., M, the formula for normalization is: A represents the source component matrix before normalization, and B represents the source component matrix after normalization.

[0032] Furthermore, the processing of wavelet threshold filtering includes:

[0033] Perform wavelet decomposition on the low-frequency noise signal subcomponent to obtain a wavelet signal q,

[0034] Set a threshold for each layer of wavelet signal, and retain the part of each layer of wavelet signal that exceeds the threshold, which is denoted as p;

[0035] Then subtract the retained part p from the wavelet signal q to get a clean wavelet signal;

[0036] Finally, the clean wavelet signal is reconstructed to obtain the sub-components after removing the low-frequency noise signal.

[0037] The beneficial effects of the present invention are that the present invention can realize the measurement and separation of low-frequency noise signals, can improve the accuracy of noise signal removal, and can reduce the loss of effective signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below with reference to the accompanying drawings and examples.

[0039] Figure 1 It is a flow chart of the signal noise detection method of the present invention.

[0040] Figure 2 The figure is a flow chart of the signal noise detection method according to the present invention according to the priority processing.

[0041] Figure 3 It is a flow chart of low-frequency noise signal positioning of the present invention.

[0042] Figure 4It is a comparison chart of the correlation coefficients of Example 1, Comparative Examples 1 and 2.

[0043] Figure 5 It is a comparison diagram of the relative root mean square errors of Example 1 and Comparative Examples 1 and 2.

[0044] Figure 6 It is a comparison chart of the interference removal processing effects of Example 2, Comparative Examples 3 and 4.

[0045] Figure 7 It is a schematic diagram of the signal time window classification detection of the present invention.

[0046] Figure 8 It is a flow chart of the generalized eigendecomposition method of the present invention.

[0047] Figure 9 It is a flow chart of multi-feature joint detection of the present invention.

[0048] Figure 10 It is a flow chart of the multi-feature classification detection of the present invention.

[0049] Figure 11 This is a result diagram of the noise signal removed by using the generalized eigendecomposition method of the present invention.

[0050] Figure 12 This is the result of removing noise from the signal using the existing technology.

[0051] Figure 13 This is a comparison chart of the results of joint feature detection and single feature detection in identifying noise signals. DETAILED DESCRIPTION

[0052] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0053] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply 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, features defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0054] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0055] like Figures 1 to 2 As shown, the signal noise detection method of the present invention includes the following steps: S1, dividing the signal time window; S2, classifying the signal time window into a noise signal time window and a normal signal time window; S3, classifying the noise signal in the noise signal time window into categories; S4, constructing different signal processing models for the noise signal time window according to the noise signal category to separate the corresponding noise signal. It should be noted that dividing the signal into multiple signal time windows and then classifying each signal time window can improve detection efficiency. After detecting the noise signal time window, the noise signal category contained in the noise signal time window can be obtained. Different categories can be separated by constructing different signal processing models, which is more targeted.

[0056] For example, the signal processing model includes: obtaining a time decomposition matrix V and a spatial decomposition matrix U based on the noise signal time window, where the time decomposition matrix V is a K×K matrix, where K is the number of time-domain sampling points; and the spatial decomposition matrix U is an M×M matrix, where M is the number of channels; extracting the first M columns from the time decomposition matrix V to form a source component matrix A, where M is the number of channels. Locating the position of the noise signal in the source component matrix A; determining whether the position of the noise signal in the source component matrix A is a null value. When the position of the noise signal is non-null, sequentially separating the noise signal and reconstructing the source to continue iteration; terminating the iteration when the position of the noise signal is a null value. Source reconstruction includes: outputting a matrix V_hat after filtering the noise signal; and performing an inverse source decomposition operation on the spatial decomposition matrix U and the matrix V_hat to obtain a signal S_hat. Of course, the signal processing model in this case can also be non-iterative, that is, separating the noise signal once.

[0057] It should be noted that the time decomposition matrix V is the decomposition of the signal in the time dimension, and each column in the time decomposition matrix V can reflect the time domain distribution information of each sub-component. The spatial decomposition matrix U is the decomposition of the signal in the spatial distribution dimension of the electrode, and each column can reflect the spatial domain distribution information of each sub-component. The time decomposition matrix V and the spatial decomposition matrix U of the noise signal time window can be obtained by singular value decomposition (SVD) or eigendecomposition (ED), which is not limited here. The singular value decomposition method (SVD) includes: decomposing the multi-channel signal matrix S into: S = U∑VT , S∈R M×K is a signal matrix with M channels and K time samples, U∈R M×M is a left singular matrix, ∑∈R M×K is the singular value matrix, V∈R K×K Is a right singular matrix. The left singular U matrix is the decomposition of the signal matrix S in the spatial dimension. Each column can reflect the spatial distribution of each sub-component. The right singular matrix V matrix is the decomposition of the signal matrix S in the time dimension. Each column of V can reflect the time domain distribution of each sub-component. The eigendecomposition method (ED) includes: M×K Perform SVD to obtain the left singular matrix U and the right singular matrix V, which can be obtained by the following ED equivalent: Calculate the spatial covariance matrix Rs of S = S*S T / K, perform ED on Rs: Rs=U∑U T , U represents the characteristic matrix of Rs, which is the left singular matrix of S; calculate the time domain covariance matrix of S Rt = S T *S / M, ED for Rt: Rt=V∑V T , V represents the characteristic matrix of Rt, which is the right singular matrix of S. Before locating the position of the noise signal, the present invention first constructs the source signal (i.e., the source component matrix A), A = [v1, v2, ..., v M ],v1~v M Represents the first M columns of the time decomposition matrix V.

[0058] It should be noted that the noise signal categories within the noise signal measurement window include: non-biological noise signals and biological noise signals within the noise signal measurement. To improve the accuracy of noise signal removal, the present invention employs different measurement and separation methods for different types of noise signals. Since biological and non-biological noise signals have different sources and influence the signal to varying degrees, the presence of multiple types of noise signals can affect signal detection, especially when very high-energy noise signals are present. Therefore, when separating noise signals, the present invention prioritizes or simultaneously removes certain types of noise signals, facilitating accurate separation between signals and interference sources. Non-biological and biological noise signals are separated based on, for example, but not limited to, energy (signal-to-noise ratio) to determine the source of the noise, i.e., the noise signal with the highest energy is processed first. When the noise signal energy of the non-biological noise signal far exceeds that of the biological noise signal, the non-biological noise signal is processed first, followed by the biological noise signal. The priority order can be set as follows: non-biological noise signal > biological noise signal. When the noise energy of abiogenic noise signals is significantly lower than that of biogenic noise signals, the biogenic noise signals are processed first, followed by the abiogenic noise signals. A priority order can be set: abiogenic noise signals < biogenic noise signals. Different signal processing models are then run sequentially based on the priority levels to gradually separate noise signals of different levels. When the noise energy of abiogenic noise signals is similar to that of biogenic noise signals, all noise types can be removed simultaneously or randomly.

[0059] As an optional implementation of a non-biological noise signal.

[0060] In this case, non-biological noise signals include, but are not limited to: power frequency noise, environmental electromagnetic interference, and step noise. Optionally, any type of noise signal can be removed simultaneously, randomly, or in order of priority. Preferably, based on the spatial decomposition matrix, power frequency noise is usually unavoidable and distributed across all channels, so it is usually removed first. Environmental electromagnetic interference is also distributed across all channels, so it is removed second-best. Step noise usually occurs in a few channels with large amplitude variations, so it is processed last. Therefore, the priority order of non-biological noise signals can be set according to the universality of the existence of noise signals as follows: power frequency noise > environmental electromagnetic interference > step noise.

[0061] (1) The processing of power frequency noise and environmental electromagnetic interference is as follows:

[0062] For example, measuring power frequency noise and environmental electromagnetic interference involves locating the power frequency noise and environmental electromagnetic interference signals using spatial and time-frequency characteristics. Separating the power frequency noise and environmental electromagnetic interference involves filtering the source components containing the power frequency noise and environmental electromagnetic interference at specific frequencies to isolate the power frequency noise and environmental electromagnetic noise signals. Both power frequency noise and environmental electromagnetic interference affect signals in all channels. Therefore, the variance of the spatial decomposition vectors corresponding to these two signals is almost zero. Furthermore, the frequency spectrum of the temporal decomposition matrix corresponding to these two signals exhibits peaks at specific frequencies. Therefore, the spatial and time-frequency characteristics can be used to locate the power frequency noise and environmental electromagnetic interference signals. Furthermore, filtering at specific frequencies can separate the power frequency and environmental electromagnetic noise signals. For example, the specific frequency of power frequency noise is 50Hz or 60Hz, while the specific frequency of environmental electromagnetic interference is generally set based on the electromagnetic frequency in the acquisition environment.

[0063] (2) The processing of step noise is as follows:

[0064] For example, measuring step noise involves taking a first-order difference on each column of the source component matrix A and determining the location of the step noise signal based on the absolute value of the difference. Separating step noise involves applying wavelet threshold filtering to the components corresponding to the step noise signal to isolate the step noise signal. Step noise often causes transient amplitude changes in certain channels. Therefore, taking a first-order difference on each column of the source component matrix A of the signal and finding the location where the absolute value of the difference significantly exceeds that of other columns indicates the location of the step noise.

[0065] As an optional implementation of a biological noise signal.

[0066] In this case, the biogenic noise signal includes at least one of a high-frequency noise signal, a low-frequency noise signal, and a periodic noise signal. Biogenic signals generally include physiological signals such as electromyography, electrooculography, electrokinesis, electrocardiography, and electroencephalography. Since electromyography is a high-frequency signal, electrooculography and electrokinesis are low-frequency signals, and electrocardiography is a periodic signal, in this case, different biogenic signals are selected as reference signals, and the types of biogenic noise signals included are different. If electromyography is used as the reference signal, its biogenic noise signal includes at least one of high-frequency noise signals, low-frequency noise signals, and periodic noise signals other than electromyography; and so on. For non-electroencephalography as the reference signal, the corresponding signal can be directly obtained based on the measurement or separation method of a single biogenic noise signal. Alternatively, the biogenic noise signal positioning or separation method can be used to remove biogenic noise signals other than the reference signal, and the reference signal can be indirectly obtained to achieve the denoising effect.

[0067] If EEG is used as the reference signal, high-frequency noise signals such as but not limited to electromyographic signals, low-frequency noise signals such as but not limited to electrooculographic signals and motor signals, and periodic noise signals such as but not limited to electrocardiographic signals; run the corresponding signal processing model simultaneously, randomly, or in order of priority to remove any type of noise signal. For example, since high-frequency noise signals have a wide frequency distribution and a wide spatial distribution, they are likely to cause abnormal recognition of other noise signals, so high-frequency noise signals are processed first; low-frequency noise signals have a large amplitude and obvious spatial characteristics, which are relatively easy to locate, and are ranked after high-frequency noise signals; finally, periodic noise signals are removed. Therefore, the priority order of biological noise signals can be: high-frequency noise signals > low-frequency noise signals > periodic noise signals. The specific processing of biological noise signals of EEG includes:

[0068] (1) The processing of high-frequency noise signals is as follows:

[0069] The measurement of high-frequency noise signals includes: calculating the energy characteristics of each column signal in the source component matrix A, calculating the high-frequency and low-frequency energy ratio r of each column signal hl , if the high-frequency and low-frequency energy ratio r hl > the set energy ratio threshold r0, the corresponding column signal is marked as a high-frequency noise signal subcomponent, and the set of all high-frequency noise signal subcomponent positions is set as the position L of the high-frequency noise signal. h The energy characteristic is, for example, power spectral density psd(m)=pwelch(v m ), the high-frequency and low-frequency energy ratio of each column signal:

[0070] Where m = 1, 2, ..., M, h1 and h2 represent the lower and upper bounds of the high frequency band, l1 and l2 represent the lower and upper bounds of the low frequency band. hl >r0, then the column signal is first marked as a high-frequency noise signal subcomponent, and the position of the high-frequency noise signal is L h That is the set of all high-frequency noise signal sub-component positions. Get the position L of the high-frequency noise signal h Afterwards, the separation of high frequency noise signals includes: h The high-frequency interference sub-components on the image are filtered or CCA (canonical correlation analysis) is performed to separate the high-frequency noise signal.

[0071] (2) The processing of low-frequency noise signals is as follows:

[0072] The measurement of low-frequency noise signals includes: standardizing the source component matrix A to obtain the source component matrix B; performing peak statistics on each column signal in the source component matrix B to obtain the number of peaks of each column signal; calculating the peak frequency p of each column signal in unit time.m1 ; If the peak frequency p m1 > the set frequency threshold p0, the corresponding column signal is marked as a low-frequency noise signal subcomponent, and the time position L of the low-frequency noise signal subcomponent is recorded. v ; Obtain the spatial decomposition matrix U based on the noise signal time window, where the spatial decomposition matrix U is an M×M matrix, where M is the number of channels; Map each column signal of the spatial decomposition matrix U and match it with the spatial position template to obtain the matching coefficient; If the matching coefficient is greater than the set matching threshold, the spatial position L of the low-frequency noise signal subcomponent is output u According to the formula L l =merge(L v ,L u ) Determine the location of the low-frequency noise signal L l .

[0073] It should be noted that the standardization of the source component matrix A is based on the mean μ of the samples. m =mean(v m ) and variance σ m =std(v m ), m = 1, 2, ..., M, M represents the number of channels, and the formula for normalization is: K represents the number of time samples, A represents the source component matrix before normalization, and B represents the source component matrix after normalization. The method has high generalization after normalization, and the fixed threshold is adapted to different individuals. Peak detection is performed on each column signal of the source component matrix B. Peak detection is performed based on three features: peak height, half-peak width, and peak-to-peak distance. After obtaining the number of peaks in each column signal, the peak frequency p that appears per unit time is calculated. m1 , if the peak frequency p of a column of signals m1 > p0, then the column signal is considered to contain a low-frequency noise signal, and the column signal is marked as a low-frequency noise signal sub-component, and the time position L of the low-frequency noise signal sub-component is recorded. v The measurement of low-frequency noise signals also requires the joint spatial position for final confirmation.

[0074] The separation of low frequency noise signal includes: l The low-frequency noise signal subcomponents on the image are subjected to wavelet threshold filtering to separate the low-frequency noise signal. The wavelet threshold filtering process includes: performing wavelet decomposition on the low-frequency noise signal subcomponents to obtain wavelet signals q, setting a threshold for each layer of wavelet signals, retaining the portion of each wavelet signal that exceeds the threshold (denoted as p, which is the noise signal), and then subtracting the retained portion p from the wavelet signal q to obtain a clean wavelet signal. Finally, the clean wavelet signal is reconstructed to obtain the subcomponents after removing the low-frequency noise signal.

[0075] like Figure 3 As shown, taking the electrooculogram noise signal as an example, the time position L of the electrooculogram noise signal subcomponent is obtained. v After that, we need to obtain the spatial position L of the electrooculogram noise signal u First, we need to obtain the spatial position template related to the electrooculogram. The spatial position template includes the blink template, the vertical eye movement template, and the horizontal eye movement template. The blink template, the vertical eye movement template, and the horizontal eye movement template can be obtained by performing spatial blink position detection, spatial vertical eye movement position detection, and spatial horizontal eye movement position detection on the signal respectively. m Map the channel position to the brain topography space, and then match it with the three spatial position templates to obtain three matching coefficients BL(m), VE(m), and HE(m). The number of each matching coefficient is M (the same as the number of channels). Set the matching threshold for each matching coefficient and record the u corresponding to the matching coefficient greater than the matching threshold. m The final output spatial position has three: spatial blink position L u1 , spatial vertical eye movement position L u2 , spatial horizontal eye movement position L u3 Finally, according to the formula L l =merge(L v ,L u1 ,L u2 ,L u3 )=L v ∩(L u1 ∪L u2 ∪L u3 ) determines the final location of the electrooculogram noise signal. The method of jointly determining the location of the low-frequency noise signal in the time domain and spatial domain can constrain the number of low-frequency noise signal locations and prevent signal loss caused by excessive subsequent processing components.

[0076] (3) The processing of periodic noise signals is as follows:

[0077] The measurement of periodic noise signals includes: standardizing the source component matrix A to obtain the source component matrix C; performing peak statistics on each column signal in the source component matrix C to obtain the number of peaks of each column signal; calculating the peak frequency p of each column signal in unit time. m2 ; If the peak frequency p m2 > the set frequency threshold p0, the corresponding column signal is marked as a periodic noise signal subcomponent, and the position of the periodic noise signal subcomponent is set to the position L of the periodic noise signal z The measurement method of the periodic noise signal is the same as the measurement method of the time position of the low-frequency noise signal, and will not be repeated here.

[0078] The separation of periodic noise signals includes: using template matching method to z The periodic interference sub-components on the signal are processed to separate the periodic noise signal. The periodic noise signal is, for example, an electrocardiogram noise signal. The processing process of the template matching method includes: detecting the periodic interference sub-component through a QRS wave detection algorithm, determining the R wave position in the sub-component, extracting all QRS waves in the signal according to the R wave position, iteratively clustering all QRS wave signals to obtain several periodic noise signal templates, and then matching the periodic interference sub-components with the periodic noise signal templates one by one, recording the periodic noise signal template X with the highest matching degree, and subtracting the periodic noise signal template X from the sub-component to obtain a clean sub-component.

[0079] The following is a comparative explanation through specific embodiments.

[0080] Example 1: This method is used to remove noise signals from simulation data, where the simulation data includes electrooculogram noise signals and myoelectric artifacts with different signal-to-noise ratios.

[0081] Comparative Example 1: The existing MARA (Multiple Artifact Rejection Algorithm) method is used to remove noise signals from simulation data.

[0082] Comparative Example 2: Using the existing ASR (Artifact Subspace Reconstruction) method to remove noise signals from simulation data.

[0083] Example 2: This method is used to remove noise signals from real EEG data containing electrooculogram noise signals.

[0084] Comparative Example 3: Using the existing MARA method to remove noise signals from real EEG data.

[0085] Comparative Example 4: Using the existing ASR method to remove noise signals from real EEG data.

[0086] The evaluation indicators for the processing effect of simulation data are: the correlation coefficient and relative root mean square error between the data after noise signal removal and the data without noise signal. The higher the correlation coefficient, the better the effect of removing artifacts, and the smaller the relative root mean square error, the better the effect of surface rolling noise signal removal.

[0087] The evaluation index for the processing effect of real data is: the waveform comparison between the original EEG data and the EEG data after removing the noise signal.

[0088] Figure 4(a) is a comparison chart of the correlation coefficients of Example 1, Comparative Example 1, and Comparative Example 2 in the strong eye electrooculogram channel. Figure 4 (b) is a comparison chart of the correlation coefficients of Example 1, Comparative Example 1, and Comparative Example 2 in the weak eye electrooculogram channel. Figure 4 (c) is a comparison chart of the correlation coefficients of Example 1, Comparative Example 1, and Comparative Example 2 in the electromyographic channels. Figure 5 (a) is a comparison diagram of the relative root mean square errors of Example 1, Comparative Example 1, and Comparative Example 2 in the strong eye electrooculogram channel. Figure 5 (b) is a comparison diagram of the relative root mean square errors of Example 1, Comparative Example 1, and Comparative Example 2 in the weak eye electrical channel. Figure 5 (c) is a comparison chart of the relative root mean square error of Example 1, Comparative Example 1, and Comparative Example 2 in the electromyographic channel. The number "1" on the horizontal axis represents Example 1, "2" represents Comparative Example 1, and "3" represents Comparative Example 2. Figure 4 and Figure 5 The results show that the correlation coefficient of this method is higher than that of the existing technology, and the relative root mean square error of this method is lower than that of the existing technology. Therefore, when removing multiple types of noise signals, this method can not only more accurately remove different types of noise signals, but also reduce the loss of effective signals when removing noise signals.

[0089] Figure 6 It is a waveform comparison of the original EEG signal, Example 2, Comparative Example 3, and Comparative Example 4. The horizontal axis is time, and the vertical axis is amplitude. It can be seen from the figure that the original signal contains multiple peaks (i.e., noise signals). After processing using this method and the two existing technologies, the peaks are filtered out. Compared with the MARA method, when removing noise signals, this method has less loss in signal details and less impact on the effective signal. Compared with the ASR method, this method still has some noise signals in the waveform after processing by the ASR method, and the removal effect is poor, while the waveform after removing the noise signal by this method is smoother.

[0090] For example, signal time windows can be detected using a time window detection method, which includes at least one of generalized feature decomposition, multi-feature classification, and multi-feature joint detection. This method is applicable to both primary and secondary signal windows. When secondary signal windows exist, they are typically labeled using the time window detection method. The labeled results of multiple secondary signal windows are then combined to determine the primary signal window's category, classifying it as a sparse noise signal window, a dense noise signal window, or a normal signal window.

[0091] like Figure 8The generalized eigendecomposition method shown includes: calculating the characteristic matrices of the signal time window and the background signal respectively; calculating the maximum generalized eigenvalue between the covariance matrix of the signal time window and the covariance matrix of the background signal; if the maximum generalized eigenvalue is greater than the set threshold, the signal time window is marked as a noise signal time window; otherwise, it is marked as a normal signal time window. It should be noted that the background signal is the EEG baseline or the full-band EEG signal, and the generalized eigenvalue can reflect the difference between the two matrices. The category of the signal time window is identified by the size of the generalized eigenvalue. For example, let R1 be the covariance matrix of the current signal time window, R2 is the covariance matrix after high-pass filtering of the current signal time window, and R b is the covariance matrix of the background signal, the maximum generalized eigenvalue α1 between R2 and R1 represents the myoelectric noise signal strength of the current signal time window, R1 and R b The maximum generalized eigenvalue α2 represents the noise signal strength of the current signal time window (regardless of the noise signal type). If α1 is greater than the set threshold, the signal time window is marked as an EMG noise signal time window; otherwise, it is a normal signal time window. If α2 is greater than the set threshold, the signal time window is marked as a noise signal time window; otherwise, it is a normal signal time window.

[0092] For example, Figure 11 This is the denoising signal processing effect diagram after the generalized feature decomposition method is used to identify the noise signal (only the noise signal is partially processed). Figure 12 This is the effect diagram of using existing technology to directly remove noise from the entire signal. Figure 11 and Figure 12 It can be seen that both the present method and the prior art suppress the noise signal. However, compared with the prior art, the present method causes less loss to the effective part of the signal.

[0093] like Figure 9 As shown, the multi-feature joint detection method includes: calculating multiple eigenvalues of a signal time window; applying thresholds to each of these eigenvalues one by one; and marking the signal time window as a noise signal window when all eigenvalues meet the threshold conditions. For example, eigenvalues can be line length, zero-crossing rate, mean, etc. K eigenvalues are calculated for the signal time window, and each eigenvalue is assessed one by one. Each eigenvalue corresponds to a set threshold. When all K eigenvalues meet their respective set thresholds, the signal time window is judged to be a noise signal window; otherwise, it is a normal signal window. For example, for a signal time window, two eigenvalues, line length and zero-crossing rate, are calculated. If the line length feature meets the set threshold, the signal time window is first marked as a candidate interference. If the zero-crossing rate feature also meets the set threshold, the signal time window is finally marked as a noise signal window. This improves the accuracy of noise signal detection and prevents misjudgments.

[0094] like Figure 13As shown in the figure, it is a comparison chart of the detection results of the multi-feature joint detection method and the single-feature detection method. The abscissa represents the time window, and the ordinate represents the channel. In the order from top to bottom, Figure 13 the first small figure in Figure 13 represents the original signal. The areas selected by the black frames (from left only) are electromyogram noise signals, normal electroencephalogram, epileptic electroencephalogram, and electromyogram noise signals respectively; the second small figure represents the retrieval result obtained only by using the line length feature, and the third small figure represents the detection result obtained by combining the two features of line length and zero-crossing rate. The white bright spots in the second and third small figures indicate the positions determined as electromyogram noise signals. It can be seen from this that there are electromyogram noise signals, normal signals, and epileptic signals in the original signal. Using the single-feature detection method, useful epileptic signals are also determined as noise signals, which will lead to inaccurate early warnings for epilepsy in the follow-up. And this method can accurately identify electromyogram noise signals and will not misjudge epileptic signals as electromyogram signals.

[0095] Such as Figure 11 shown, the multi-feature classification method includes: calculating multiple feature values of the signal time window, and inputting the feature values into a classifier for classification to output the time windows of noise signals and normal signals. The feature values are, for example, line length, zero-crossing rate, mean value, etc. After calculating the feature values of the signal time window, input the feature values into the classifier, and the classifier can output the classification result (noise signal or normal signal). Classifying the data of M channels in the signal time window one by one can obtain the spatio-temporal classification result of the signal time window.

[0096] Such as Figure 7 shown, measuring the noise signal categories in the noise signal time window also includes: dividing the noise signal time window into sparse noise signal time window and dense noise signal time window according to the spatio-temporal sparsity of the noise signal. Dividing the signal time window includes at least two levels of division, that is, a first-level signal time window is composed of multiple second-level signal time windows. Classifying and detecting the signal time window also includes: jointly judging the first-level signal time window based on the detection results of the second-level signal time windows to divide the first-level signal time window into sparse noise signal time window, dense noise signal time window, and normal signal time window. Jointly judging the first-level signal time window based on the detection results of the second-level signal time windows includes: detecting the second-level signal time windows and dividing them into second-level interference time windows and second-level normal time windows; calculating the proportion p of the second-level interference time windows in the first-level signal time window, that is, when the proportion p = 0, marking the first-level signal time window as a normal signal time window; when 0 < p < the proportion threshold, marking the first-level signal time window as a sparse noise signal time window; when p ≥ the proportion threshold, marking the first-level signal time window as a dense noise signal time window.

[0097] In other words, in addition to classifying noise signals according to the type of noise signals, the present invention can also classify the time windows of noise signals according to the sparsity degree of the noise signals. First, the signal time window is divided into two levels, namely the secondary signal time window and the primary signal time window. The primary signal time window is composed of multiple secondary signal time windows. That is, the window width of the primary signal time window is greater than that of the secondary signal time window, and the amount of data contained in the primary signal time window is also more than that of the secondary signal time window. Therefore, when the present invention performs data processing, it first detects the secondary signal time window to obtain the detection result of each secondary signal time window (divided into secondary interference time window and secondary normal time window). In this way, it can also reflect how many secondary interference time windows are contained in the primary signal time window, and the proportion p = the number of secondary interference time windows / the number of secondary signal time windows. If p = 0, it indicates that there is no secondary interference time window, and this primary signal time window is a normal signal time window and does not need further processing. If 0 < p < the proportion threshold, it indicates that there are a small number of secondary interference time windows, and this primary signal time window is a sparse noise signal time window. At this time, when removing the noise signal, only the secondary interference time window needs to be processed, and there is no need to process the entire primary signal time window. If the proportion p ≥ the proportion threshold, it indicates that there are more secondary interference time windows, and this primary signal time window is a dense noise signal time window. At this time, when removing the noise signal, the entire primary signal time window needs to be processed.

[0098] When the primary signal time window is a sparse noise signal time window, different signal processing models are only constructed for the secondary interference time windows in the primary signal time window. When the primary signal time window is a dense noise signal time window, a different signal processing model is constructed for the primary signal time window. That is, for the primary signal time window of "sparse noise signal", only the secondary interference time window is processed. In this way, it can avoid processing non-noise signal segments, prevent the loss of effective signals, and reduce the amount of data to be processed, thereby improving the processing efficiency. For the primary signal time window of "dense noise signal", the entire primary signal time window is processed, which has higher operation efficiency and can also prevent missed processing.

[0099] It should be noted that the set thresholds involved in this case can be set according to requirements, historical data or technical experience, and are not specifically limited in this case.

[0100] The signal noise detection method of the present invention uses a spatio-temporal domain joint method to measure and separate multi-type noise signals of the signal, which can improve the accuracy of noise signal removal and reduce the loss of effective signals. Moreover, for different types of noise signals, different measurement methods and separation methods are adopted. Compared with using the same processing method, this method is more targeted and the recognition result will be more accurate.

[0101] The present invention also provides a signal detection model for a signal noise detection method, comprising: a signal time window division model for dividing the signal time window; a signal time window detection model for classifying and detecting the signal time window to obtain noise signal time windows and normal signal time windows; a classification model for measuring the noise signal category within the noise signal time window; and a signal processing model for separating different noise signal categories based on the noise signal category. For a description of the signal noise detection method, please refer to the section on the signal noise detection method and will not be repeated here.

[0102] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the signal noise detection method. The computer-readable storage medium may be located on at least one of a plurality of network servers in a computer network. The storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a removable hard drive, a magnetic disk, or an optical disk.

[0103] The present invention also provides a signal window detection model based on the generalized eigendecomposition method, comprising: calculating the characteristic matrices of the signal window and the background signal respectively; calculating the maximum generalized eigenvalue between the characteristic matrices of the signal window and the background signal; if the maximum generalized eigenvalue is greater than a set threshold, marking the signal window as a noise signal window; otherwise, marking it as a normal signal window. A detailed description is provided in the relevant section of the signal noise detection method and will not be repeated here.

[0104] The present invention also provides a signal window detection model based on a multi-feature classification method, comprising: calculating multiple eigenvalues of a signal window; applying thresholds to each of the multiple eigenvalues; and marking the signal window as a noise signal window when all the eigenvalues of the signal window meet the thresholds. A detailed description of this method is provided in the relevant section of the signal noise detection method and will not be repeated here.

[0105] The present invention also provides a signal time window detection model based on a multi-feature joint detection method, which includes calculating multiple eigenvalues of a signal time window, inputting the eigenvalues into a classifier for classification, and outputting a noise signal time window and a normal signal time window. A detailed description is provided in the relevant section of the signal noise detection method and will not be repeated here.

[0106] The present invention also provides a high-frequency noise signal detection model, including: dividing the signal time window; classifying and detecting the signal time window into a noise signal time window and a normal signal time window; constructing a signal processing model for the noise signal time window to separate the high-frequency noise signal. The signal processing model includes: obtaining a time decomposition matrix V based on the noise signal time window, wherein the time decomposition matrix V is a K×K matrix, and K is the number of time domain sampling points; extracting the first M columns from the time decomposition matrix V to form a source component matrix A, wherein M is the number of channels; locating the position of the high-frequency noise signal in the source component matrix A; determining whether the position of the high-frequency noise signal in the source component matrix A is a null value; when the position of the high-frequency noise signal is a non-null value, separating the high-frequency noise signal and reconstructing the source in sequence to continue iteration; terminating the iteration when the position of the high-frequency noise signal is a null value. Locating the position of the high-frequency noise signal in the source component matrix A includes: calculating the energy characteristics of each column signal in the source component matrix A; calculating the high-frequency and low-frequency energy ratio r of each column signal. hl ; If the high-frequency and low-frequency energy ratio r hl > the set energy ratio threshold r0, the corresponding column signal is marked as a high-frequency noise signal subcomponent, and the set of all high-frequency noise signal subcomponent positions is set as the position L of the high-frequency noise signal. h . Separation of high frequency noise signal includes: h The high-frequency interference components on the signal are low-pass filtered or CCA filtered to separate the high-frequency noise signal. For details, please refer to the relevant part of the signal noise detection method, which will not be repeated here.

[0107] The present invention also provides a low-frequency noise signal detection model, comprising: dividing a signal time window; classifying and detecting the signal time window into a noise signal time window and a normal signal time window; and constructing a signal processing model for the noise signal time window to separate the low-frequency noise signal. The signal processing model comprises: obtaining a spatial decomposition matrix U and a temporal decomposition matrix V based on the noise signal time window, wherein the spatial decomposition matrix U is an M×M matrix, where M is the number of channels, and the temporal decomposition matrix V is a K×K matrix, where K is the number of time domain sampling points; extracting the first M columns from the temporal decomposition matrix V to form a source component matrix A, where M is the number of channels; locating the position of the low-frequency noise signal in the source component matrix A; determining whether the position of the low-frequency noise signal in the source component matrix A is a null value; when the position of the low-frequency noise signal is a non-null value, sequentially separating the low-frequency noise signal and reconstructing the source to continue iteration; and terminating the iteration when the position of the low-frequency noise signal is a null value. The location of the low-frequency noise signal in the source component matrix A includes: standardizing the source component matrix A to obtain the source component matrix B; performing peak statistics on each column signal in the source component matrix B to obtain the number of peaks of each column signal; calculating the peak frequency p of each column signal in unit time. m1 ; If the peak frequency p m1> the set frequency threshold p0, the corresponding column signal is marked as a low-frequency noise signal subcomponent, and the time position L of the low-frequency noise signal subcomponent is recorded. v ; Match each column signal of the spatial decomposition matrix U with the residual spatial position template after mapping to obtain the matching coefficient; if the matching coefficient is greater than the set matching threshold, the spatial position L of the low-frequency noise signal subcomponent is output u According to the formula L l =merge(L v ,L u ) Determine the location of the low-frequency noise signal L l . Separating low frequency noise signals includes: l The low-frequency interference subcomponents on the signal are subjected to wavelet threshold filtering to separate the low-frequency noise signal. For a detailed description, please refer to the relevant part of the signal noise detection method, which will not be repeated here.

[0108] The present invention also provides a periodic noise signal detection model, including: dividing the signal time window; classifying and detecting the signal time window into a noise signal time window and a normal signal time window; constructing a signal processing model for the noise signal time window to separate the periodic noise signal. The signal processing model includes: obtaining a time decomposition matrix V based on the noise signal time window, wherein the time decomposition matrix V is a K×K matrix, and K is the number of time domain sampling points; extracting the first M columns from the time decomposition matrix V to form a source component matrix A, wherein M is the number of channels; locating the position of the periodic noise signal in the source component matrix A; judging whether the position of the periodic noise signal in the source component matrix A is a null value; when the position of the periodic noise signal is a non-null value, separating the periodic noise signal and reconstructing the source in sequence to continue iteration; when the position of the periodic noise signal is a null value, terminating the iteration. Locating the position of the periodic noise signal in the source component matrix A includes: standardizing the source component matrix A to obtain a source component matrix C; performing peak statistics on each column signal in the source component matrix C to obtain the number of peak values of each column signal; calculating the peak frequency p of each column signal occurring per unit time. m2 ; If the peak frequency p m2 > the set frequency threshold p0, the corresponding column signal is marked as a periodic noise signal subcomponent, and the position of the periodic noise signal subcomponent is set to the position L of the periodic noise signal z The separation of periodic noise signals includes: using template matching method to z The periodic interference subcomponents on the signal are processed to separate the periodic noise signal. For a detailed description, please refer to the relevant part of the signal noise detection method, which will not be repeated here.

[0109] The present invention also provides a multi-level detection model for signal time windows, including: at least two levels of dividing signal time windows, that is, a first-level signal time window is composed of multiple second-level signal time windows; using the time window detection method to detect the second-level signal time windows, which are divided into second-level interference time windows and second-level normal time windows; based on the detection results of the second-level signal time windows, jointly judge the first-level signal time windows to divide the first-level signal time windows into sparse noise signal time windows, dense noise signal time windows, and normal signal time windows. The time window detection method is at least one of the generalized eigenvalue decomposition algorithm, multi-feature classification algorithm, and multi-feature joint detection algorithm. The generalized eigenvalue decomposition method includes: respectively calculating the feature matrices of the signal time window and the background signal; calculating the maximum generalized eigenvalue between the feature matrix of the signal time window and the feature matrix of the background signal; if the maximum generalized eigenvalue > the set threshold, mark the signal time window as a noise signal time window; otherwise, mark it as a normal signal time window. The multi-feature joint detection method includes: calculating multiple eigenvalues of the signal time window; performing threshold judgment on each of the multiple eigenvalues one by one. When all the multiple eigenvalues of the signal time window meet the threshold conditions, mark the signal time window as a noise signal time window. The multi-feature classification method includes: calculating multiple eigenvalues of the signal time window and inputting the eigenvalues into a classifier for classification, outputting noise signal time windows and normal signal time windows. Jointly judging the first-level signal time windows based on the detection results of the second-level signal time windows includes: calculating the proportion p of the second-level interference time windows in the first-level signal time window; when the proportion p = 0, mark the first-level signal time window as a normal signal time window; when 0 < p < the proportion threshold, mark the first-level signal time window as a sparse noise signal time window; when the proportion p ≥ the proportion threshold, mark the first-level signal time window as a dense noise signal time window. For specific elaboration, please refer to the relevant part of the signal noise detection method, which will not be elaborated here.

[0110] In summary, the signal noise detection method, signal detection model, and readable storage medium of the present invention use a spatio-temporal domain joint method to measure and separate multiple types of noise signals from signals, which can improve the accuracy of noise signal removal and reduce the loss of effective signals. Moreover, for different types of noise signals, different measurement methods and separation methods are adopted. Compared with using the same processing method, this method is more targeted and the recognition result will be more accurate.

[0111] Taking the above ideal embodiments based on the present invention as inspiration, through the above description, relevant staff can make various changes and modifications completely within the scope of not deviating from the technical idea of this invention. The technical scope of this invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A low-frequency noise signal detection method, characterized in that: include: Signal time window division model, divides the signal time window; Signal time window detection model, which divides the signal time window into noise signal time window and normal signal time window; A signal processing model is used to separate the low-frequency noise signal in the noise signal window; The signal processing model includes: Acquire a spatial decomposition matrix U and a temporal decomposition matrix V based on the noise signal time window; Extract the first M columns from the time decomposition matrix V to form the source component matrix A, where M is the number of channels; Locate the position of the low-frequency noise signal in the source component matrix A; Determine whether the position of the low-frequency noise signal in the source component matrix A is a null value; When the position of the low-frequency noise signal is a non-null value, separating the low-frequency noise signal and reconstructing the source are sequentially performed to continue iteration; When the position of the low-frequency noise signal is a null value, the iteration is terminated.

2. The low-frequency noise signal detection method according to claim 1, wherein: The spatial decomposition matrix U is an M×M matrix, and the temporal decomposition matrix V is a K×K matrix, where K is the number of time domain sampling points.

3. The low-frequency noise signal detection method according to claim 1, wherein: The source reconstruction includes: Output the matrix V_hat after filtering the low-frequency noise signal; An inverse source decomposition operation is performed on the spatial decomposition matrix U and the matrix V_hat to obtain a signal S_hat.

4. The low-frequency noise signal detection method according to claim 1, wherein: The position of the low-frequency noise signal in the positioning source component matrix A includes: Performing standardization on the source component matrix A to obtain a source component matrix B; Performing peak statistics on each column of signals in the source component matrix B to obtain the number of peak values of each column of signals; Calculate the peak frequency p of each column signal per unit time m1 ; If the peak frequency p m1 > the set frequency threshold p0, the corresponding column signal is marked as a low-frequency noise signal sub-component, and the time position L of the low-frequency noise signal sub-component is recorded. v ; After mapping, each column signal of the spatial decomposition matrix U is matched with the residual spatial position template to obtain the matching coefficient; If the matching coefficient is greater than the set matching threshold, the spatial position L of the low-frequency noise signal subcomponent is output. u ; According to the formula L l =merge(L v ,L u ) Determine the location of the low-frequency noise signal L l .

5. The low-frequency noise signal detection method according to claim 4, wherein: The separating of the low-frequency noise signal comprises: For position L l The low-frequency interference sub-components on the image are processed by wavelet threshold filtering to separate the low-frequency noise signal.

6. The low-frequency noise signal detection method according to claim 4, wherein: The normalization process of the source component matrix A includes: Calculate the mean μ of each column signal in the source component matrix A m =mean(v m ) and variance σ m =std(v m ), m=1, 2, ..., M, the formula for normalization is: A represents the source component matrix before normalization, and B represents the source component matrix after normalization.

7. The low-frequency noise signal detection method according to claim 5, wherein: The processing of wavelet threshold filtering includes: Perform wavelet decomposition on the low-frequency noise signal subcomponent to obtain a wavelet signal q, Set a threshold for each layer of wavelet signal, and retain the part of each layer of wavelet signal that exceeds the threshold, which is denoted as p; Then subtract the retained part p from the wavelet signal q to get a clean wavelet signal; Finally, the clean wavelet signal is reconstructed to obtain the sub-components after removing the low-frequency noise signal.

Images