CN107450730B - A method and system for slow eye movement recognition based on convolutional hybrid model - Google Patents

Abstract

The invention discloses a slow eye movement recognition method and system based on a convolution mixed model, belonging to the technical field of electrooculography, comprising using a complex-valued ICA algorithm to perform blind source separation on eye movement data of each frequency point to obtain independent The frequency domain independent components of the source signal at the corresponding frequency points; the scale compensation is performed on the independent components at each frequency point to restore the true proportional components of the independent components in the observed components; the constrained DOA algorithm is used to sort and adjust the compensated independent components ; Perform inverse short-time Fourier transform processing on the independent components of each frequency point after scale compensation and sorting to obtain the complete time signal of the multi-channel independent source in the time domain; perform wavelet decomposition on the complete time signal of the multi-channel independent source , the obtained decomposition results are compared and analyzed with the evaluation criteria of slow eye movement, and those that are consistent with the characteristics of slow eye movement are identified as slow eye movement. The present invention performs wavelet analysis on the multi-channel EOG signal in the time domain, and can quickly extract the slow eye movement from the EOG signal because there is no interference from other source signals.

Description

A method and system for slow eye movement recognition based on convolution mixture model

technical field

The invention relates to the technical field of electrooculography, in particular to a slow eye movement recognition method and system based on a convolution hybrid model.

Background technique

The visual system is the most important channel for humans to obtain external information. In the early history of experimental psychology, psychologists began to pay attention to the psychological significance of eye movement characteristics and their laws. The information processing mechanism has also become a research hotspot in contemporary psychology. The eye movement feature is the movement of the eyeball, which is closely related to the internal information processing mechanism. There are two kinds of control, exogenous and endogenous. In more cases, the eye movement will be guided by the task or purpose.

As a low-cost eye movement signal measurement technology, Electro-oculogram (EOG) is not only more accurate in measurement than traditional video methods, but also has the advantages of light weight, convenient for long-term recording, and easier measurement. Realize the advantages of wearable design and so on. Therefore, the use of EOG for eye movement signal acquisition has broad application prospects.

According to the characteristics of eye movement, the types of eye movement signals are mainly divided into three categories: fixation, saccade and follow-up eye movement. Gaze is often accompanied by three forms of subtle slow eye movements: spontaneous high-frequency eye movement microfibrillation, slow eye movement, and microsaccades. Among them, slow eye movement refers to the eye movement during the transition from wakefulness to sleep, which is called slow eye movement. During the collection of EOG data, slow eye movements existed in the collected eye movement data because the subjects were prone to fatigue.

Slow eye movement contains a lot of useful information, and has a wide range of applications in many fields such as traffic psychology and clinical medicine. For example, the driver's fatigue level can be detected by slow eye movement, and some complex clinical cases can be studied and treated by slow eye movement. But in practical applications, slow eye movements are usually intertwined with other signals, and it is difficult to extract them individually. At present, the linear regression method is used to identify slow eye movements, mainly through the ratio between the number of detected slow eye movements and the number of visual detections. However, this indicator does not clarify the performance of the algorithm, the recognition results are not ideal, and it is difficult to achieve practical purposes.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a slow eye movement recognition method and system based on a convolution hybrid model, so as to improve the accuracy of slow eye movement recognition.

In order to achieve the above objects, in the first aspect, the present invention provides a slow eye movement recognition method based on a convolutional hybrid model, including:

S1. In the frequency domain, the complex-valued ICA algorithm is used to perform blind source separation on the eye movement data of each frequency point to obtain the frequency domain independent components of each independent source signal at the corresponding frequency point;

S2. Perform scale compensation on the independent components at each frequency point to restore the true proportional components of the independent components in the observed components;

S3. The constrained DOA algorithm is used to sort and adjust the independent components after compensation, so that the independent sources on each frequency point are arranged according to the direction angle from small to large;

S4, performing short-time inverse Fourier transform processing on the independent components of each frequency point after scale compensation and sorting, to obtain a complete time signal of multi-channel independent sources in the time domain;

S5. Perform wavelet decomposition on the multi-channel independent sources in the time domain to obtain wavelet coefficients at all levels;

S6, compare and analyze the wavelet coefficients at all levels and the evaluation criteria of slow eye movement, and identify the slow eye movement if they are consistent with the slow eye movement characteristics.

Wherein, in the step S5, the generating function of wavelet decomposition adopted is db4, and the number of layers of wavelet decomposition is ten layers.

The features of the slow eye movement include: the frequency of the eye movement signal is lower than 1 Hz, the initial movement speed of the eye movement signal is approximately 0, and no artifact signal appears in the EOG signal.

Wherein, the step S2 specifically includes:

According to the separation matrix of each frequency point in the complex-valued ICA algorithm, the mixing matrix of the corresponding frequency point is obtained, wherein the separation matrix and the mixing matrix are mutually inverse matrices;

The independent components of each frequency point are compensated by the coefficients of the mixing matrix, and the independent components of each frequency point after compensation are obtained.

Wherein, the step S3 specifically includes:

a. Initialize an angle for each independent source;

b. Calculate the different rows of each frequency point through the Root-Music algorithm to obtain the estimation of each source direction, where the rows of the separation matrix correspond to different independent sources;

c. Set the proximity measure between the direction angle of each independent source and the initialization angle as ε(y, θ), and in the iterative process, determine whether the angle of each independent source is the same as the initialization angle;

d. If they are the same, execute step e, and if they are not the same, execute step f;

e. Set ε(y _j , θ _j ) to 0, and set the direction angle matrix T to calculate the adjustment matrix Q;

f. Set ε(y _j , θ _j ) to 1, and return to the iterative process to recalculate the separation matrix W.

Wherein, before the step S1, it also includes:

Collect multi-channel EOG data to obtain eye movement data in the time domain;

Perform band-pass filtering and de-average processing on the eye movement data in the time domain to obtain the processed eye movement data;

Perform short-time Fourier transform on the processed eye movement data, transform it from the time domain to the frequency domain, and obtain the eye movement data in the frequency domain.

In a second aspect, the present invention provides a slow eye movement recognition system based on a convolution hybrid model, comprising: a blind source separation module, a scale compensation module, a sorting module, a restoration module, a wavelet decomposition module and a slow eye movement connected in sequence identification module;

The blind source separation module is used to perform blind source separation on the eye movement data of each frequency point using the complex-valued ICA algorithm in the frequency domain, and obtain the frequency domain independent components of each independent source signal at the corresponding frequency point and separate the frequency domain independent components. Transfer to the scale compensation module;

The scale compensation module is used to perform scale compensation for the independent components on each frequency point, restore the true proportional components of the independent components in the observed components, and transmit the compensated independent components to the sorting module;

The sorting module is used to sort and adjust the compensated independent components by using the constrained DOA algorithm, so that the independent sources on each frequency point are arranged according to the direction angle from small to large;

The recovery module is used to perform inverse short-time Fourier transform processing on the independent components of each frequency point after scale compensation and sorting, and obtain the complete time signal of the multi-channel independent source in the time domain and the complete time signal of the multi-channel independent source. Transfer to the decomposition module;

The wavelet decomposition module is used to perform wavelet decomposition on the complete time signal of multi-channel independent sources in the time domain, obtain wavelet coefficients at all levels, and transmit the decomposition results to the slow eye movement recognition module;

The slow eye movement recognition module is used to compare and analyze the wavelet coefficients at all levels with the slow eye movement evaluation criteria, and identify the slow eye movement if they are consistent with the slow eye movement characteristics.

Compared with the prior art, the present invention has the following technical effects: the present invention adopts the complex-valued ICA algorithm to carry out blind source separation of the frequency domain observation data, and solves the problem of the inherent scale inconsistency of the ICA by restoring the true proportional component of the independent component in the observation component. At the same time, the inherent sorting fuzzy problem of ICA is solved by constrained DOA algorithm, and each independent source is separated and converted into data in the time domain. Since the time domain signals after inverse Fourier transform do not interfere with each other, in the Wavelet analysis is performed on the multi-channel EOG signal in the time domain, and slow eye movement analysis is performed on the decomposed wavelet coefficients at all levels. Since there is no interference from other source signals, the accuracy is high, the amount of calculation is small, and the EOG can be quickly obtained. Slow eye movements are extracted from the signal.

Description of drawings

Below in conjunction with the accompanying drawings, the specific embodiments of the present invention are described in detail:

Fig. 1 is a schematic diagram of the distribution of electrodes on the subject's face during EOG signal acquisition in the present invention;

2 is a schematic flowchart of a slow eye movement recognition method based on a convolution hybrid model in the present invention;

Fig. 3 is the robust saccade identification algorithm flow chart of multi-channel EOG signal in the present invention;

Fig. 4 is the basic principle diagram of blind source separation (Blind Source Separation, BSS) in the present invention;

Fig. 5 is the time-frequency domain waveform diagram of six adjacent frequency points of the present invention;

Fig. 6 is the EOG waveform diagram before and after convolution ICA separation of the present invention;

Fig. 7 is linear ICA model and convolution ICA model separation result contrast diagram in the present invention;

Fig. 8 is the average recognition rate under different methods of the present invention;

Fig. 9 is the result graph of slow eye movement experiment in the present invention;

FIG. 10 is a schematic structural diagram of a slow eye movement recognition system based on a convolution hybrid model in the present invention.

The present invention will be further described below through specific embodiments and in conjunction with the accompanying drawings.

Detailed ways

To further illustrate the features of the present invention, please refer to the following detailed description and accompanying drawings of the present invention. The attached drawings are for reference and description only, and are not intended to limit the protection scope of the present invention.

First of all, it should be noted that in the present invention, before the EOG signal is identified, the process of collecting the EOG signal is as follows:

As shown in Figure 1, electrodes were used to collect the EOG signal of the subject, and Ag/AgCl electrodes were used to collect the EOG signal. There is a lot of spatial position information, and a total of 9 electrodes were used in this acquisition. Among them, electrode V1 and electrode V2 were placed 1.5cm above and 1.5cm below the left (or right) eyeball of the subject to collect vertical EOG signal; Electrode H1 and Electrode H2 were placed 1.5cm on the left side of the subject's left eye and 1.5cm on the right side of the right eye to collect horizontal EOG signals; Electrode Fp1 and Electrode Fp2 were placed on the forehead to enhance the space Information; reference electrodes C1 and C2 are placed on the left and right breasts, respectively, and the ground electrode D is located at the center of the top of the head.

During the specific experimental collection, the subject sat in front of the screen and faced the screen. A "prepare" character appeared on the screen accompanied by a "beep" alarm sound. After 1 second, the subject could see a beep on the screen. The red arrow prompts (respectively: up arrow, down arrow, left arrow and right arrow), the arrow continues to appear on the screen for 6 seconds, during this time, the experiment requires subjects to see the arrow after seeing the arrow Eyes were rolled in the direction indicated by the arrows, and returned to the center point after seeing the observation point, during which the subject was unable to blink. Afterwards, there was a short 2-second break in which the subject could blink and relax.

As shown in FIG. 2 to FIG. 3 , the present invention discloses a slow eye movement recognition method based on a convolution hybrid model, including the following steps S1 to S6:

S1. In the frequency domain, the complex-valued ICA algorithm is used to perform blind source separation on the eye movement data of each frequency point to obtain the frequency domain independent components of each independent source signal at the corresponding frequency point;

It should be noted that in this embodiment, the multi-channel EOG data collected in the time domain is subjected to band-pass filtering and de-averaging processing, wherein the cut-off frequency of the band-pass filter is 0.01 Hz to 8 Hz, and then the window length is 256. A sliding window with a window shift of 128 performs a Short-Time Fourier Transform (STFT) on the processed eye movement data, and transforms the eye movement data in the time domain into the eye movement data in the frequency domain.

In this embodiment, band-pass filtering and de-averaging are performed on the eye movement signal in the time domain to remove the interference of signals including baseline drift, EMG, ECG, and EEG, and reduce the impact of different noise signals on the original signal. The interference of multi-channel eye movement data, thereby improving the recognition accuracy.

As shown in Figure 4, the process of blind source separation for eye movement data in the frequency domain is as follows:

1) According to the multi-channel observation data X _i (i=1,2,...N), calculate the covariance matrix R _x of the observation data, and the calculation formula of the covariance matrix is: R _x =E{(Xm _x )(Xm _x ) ^T } ^T , where X is the observation data, m _X is the mean value of the observation data, (·) ^T represents the transpose operation of the formula in the brackets, and E{·} is the expectation operation of the data in the brackets; After the covariance matrix R _x of the observation data is obtained, the observation data needs to be whitened to realize the orthogonalization of the mixture matrix, and the whitening matrix V is calculated. The calculation process is as follows:

The covariance matrix R _x is decomposed into: R _x =EDE ^T , where E is a matrix composed of the normalized orthogonal eigenvectors of R _x , and D=diag(λ ₁ ,λ ₂ ,...,λ _N ) is given by A diagonal matrix of eigenvalues corresponding to eigenvectors.

The expression form of the obtained whitening matrix is: V=D ^-12 E ^T .

2) Using the whitening matrix, whiten the observed data through the formula Z(t)=VX(t), and obtain the fourth-order cumulant of the whitening process, and use the formula N={λ; N _r |1<<r <<M} Important features whose number of calculations does not exceed M, where λ is the eigenvector, N _r represents the dimension of the observation data, M represents the number of sources, and r is an integer that does not exceed the number of sources;

3) Use the unitary matrix to jointly diagonalize the formula N={λ; N _r |1<<r<<M}, where the unitary matrix is U, and calculate the mixing matrix A by the formula A=W×U;

4) Since the mixing matrix A and the separation matrix W are mutually inverse matrices, the separation matrix W=A ^-1 , and blind source separation can be performed on the observation data at each frequency point according to the separation matrix W.

S2. Perform scale compensation on the independent components at each frequency point to restore the true proportional components of the independent components in the observed components;

Specifically, the specific process of scale compensation for independent components is as follows:

According to the separation matrix of each frequency point in the complex-valued ICA algorithm, the mixing matrix of the corresponding frequency point is obtained, wherein the separation matrix and the mixing matrix are mutually inverse matrices;

The independent components of each frequency point are compensated by the coefficients of the mixing matrix, and the independent components of each frequency point after scale compensation are obtained.

Specifically, taking the two-dimensional ICA problem as an example, define the observed signals as x ₁ , x ₂ , and the sources as s ₁ , s ₂ , then the observed signals can be expressed as:

x ₁ =a ₁₁ s ₁ +a ₁₂ s ₂ =v ₁₁ +v ₁₂ ,

x ₂ =a ₂₁ s ₁ +a ₂₂ s ₂ =v ₂₁ +v ₂₂ .

Among them, v _ij =a _ij s _ij represents the true component of the independent source s _j in the observation signal _xi , that is, the projection of the independent source s _j in the observation signal _xi , since v ₁₁ and v ₂₁ both come from the independent source s ₁ , And v ₁₁ , v ₂₁ and s ₁ are only different in magnitude. Likewise, the relationship of v ₁₂ , v ₂₂ to the independent source s ₂ is the same. Therefore, if W(f _k ) is the estimated separation matrix of a certain frequency point, the mixing moment A(f _k )=W ⁻¹ (f _k ) at the frequency point can be obtained. Then, the independent components of each frequency point can be compensated by using the obtained mixing matrix coefficients, namely:

Among them, Y _j (f _k ,τ) represents the independent component of the jth channel that has been separated before scale compensation, v _ij (f _k ,τ)=A _ij (f _k )Y _ij (f _k ,τ) represents the scaled After compensation, the true component of the jth independent component in the ith observation signal. According to the above analysis, after using the above formula to perform scale compensation on the independent components of a certain frequency point f _k , one frequency domain independent component will generate N compensated outputs, and these N compensation structures will be used for subsequent steps such as eliminating sorting ambiguity, The combination of different frequency points and the inverse transformation are processed to obtain N pure signals from the same independent source.

In practical applications, N pure signals from the same independent source can be selected for output, or N signals from the same independent source can be averaged and then output.

S3. The constrained DOA algorithm is used to process the compensated independent components, so that the independent sources on each frequency point are arranged according to the direction angle from small to large;

Specifically, the process of sorting the compensated independent components is as follows:

a. Initialize an angle for each independent source;

b. Calculate the different rows of each frequency point through the Root-Music algorithm to obtain the estimation of each source direction, where the rows of the separation matrix correspond to different independent sources;

c. Set the proximity measure between the direction angle of each independent source and the initialization angle as ε(y, θ), and in the iterative process, determine whether the angle of each independent source is the same as the initialization angle;

d. If they are the same, execute step e, and if they are not the same, execute step f;

e. Set ε(y _j , θ _j ) to 0, and set the direction angle matrix T to calculate the adjustment matrix Q;

f. Set ε(y _j , θ _j ) to 1, and return to the iterative process to recalculate the separation matrix W.

It should be noted that an angle θ _j is initialized for each independent source s _j . Due to the uncertainty of the position of the independent source, the angle of the i-th independent source is set to be smaller than the angle of the i+1-th source, and the initialized The angle r(θ) is used as a constraint. Assuming that the separation of each frequency point f is successful, the rows of the separation matrix W correspond to different independent sources, and the Root-Music algorithm is used to calculate the different rows of each frequency point, and then the estimation of each source direction can be obtained.

Here, in order to effectively compare the constrained DOA algorithm's ability to identify and sort order errors, the proximity measurement between the angle obtained by each frequency point and the initialization angle is set as ε(y _j , θ _j ), where y _j is the estimation of each source direction , compare the two angles in the iterative process, if they are not the same, that is, ε(y _j , θ _j )=1, then return to the iterative process to recalculate the separation matrix W.

If the two angles are the same, that is, ε(y _j , θ _j )=0, a direction angle matrix T needs to be set to calculate the adjustment matrix Q.

Among them, the independent sources on each frequency point f are arranged in the order of the angles from small to large, and the direction angle matrix T is set as:

In the direction angle matrix T, the diagonal line shows the angular arrangement order. After performing amplitude compensation on the signal after blind source separation, the estimated y of the source signal S can be obtained:

y=P∧S=PV,

Among them, P is the permutation matrix, ∧ is the diagonal matrix, and S is the source signal. In order to facilitate analysis, the X in y=WX is brought in with AS, and y=WAS=Ds can be obtained. Among them, W is the separation matrix, X is the observation data, A is the mixing matrix, and S is the source signal. According to the uncertainty of ICA, each row and column of matrix D must have only one non-zero element. It can be transformed into D=P∧, where P is the permutation matrix and ∧ is the diagonal matrix. P and ∧ introduce uncertainties in the ordering and magnitude of the ICA output, respectively. Therefore, in this embodiment, an adjustment matrix Q is set to perform adjustment calculation on the P matrix, so as to solve the problem of sorting uncertainty of ICA.

Further, the process of calculating the adjustment matrix Q by the direction angle matrix T is:

Q=TP ^-1 ,

At this time, if the permutation matrix P is the same as the direction angle matrix T, then the independent sources on each frequency point have been arranged in the order from small to large, so there is no need to re-adjust;

If the permutation matrix P is different from the direction angle matrix T, then the permutation matrix P is left-multiplied by the adjustment matrix Q to obtain a new permutation matrix P`;

The permutation matrix P is processed by the formula P`=QP=TP <sup>-1</sup> P=T to obtain a new permutation matrix P`, and then the direction of the independent source at the frequency point is re-obtained by the formula y=P`∧S=P`V horn. The independent sources at each frequency point f obtained at this time are arranged according to the direction angle from small to large, which solves the fuzzy sorting problem of ICA.

It should be noted that the permutation ambiguity of the ICA output is an inherent limitation of the ICA algorithm. In the blind deconvolution in the time-frequency domain, it involves the blind separation of multiple frequency points in different windows. If the ICA separation results of each frequency point are not matched, the frequency-domain independent components belonging to the same source are combined together, that is, from The sub-band signals of different sources are spliced together by mistake, which will have a great impact on the final separation effect, making the signal restored to the time domain chaotic, and then affecting the recognition result of the EOG signal. The constrained DOA algorithm is used to sort and adjust the independent components of each frequency point after amplitude compensation, so that the independent sources on each frequency point f are arranged according to the direction angle from small to large, which can effectively solve the fuzzy sorting problem of each frequency point. Improving the quality of blind source separation helps to improve the recognition rate.

S4, performing short-time inverse Fourier transform processing on the independent components of each frequency point after scale compensation and sorting, to obtain a complete time signal of multi-channel independent sources in the time domain;

It should be noted that it is used to perform short-time inverse Fourier transform under the condition that the components corresponding to different sources at each frequency point are correctly arranged and the amplitudes are recovered, and finally the obtained time-domain signals are re-intercepted and combined to obtain an estimate of the source signal.

The short-time inverse Fourier transform process is:

When calculating, invert the obtained time-frequency matrix by column to obtain time signals at different time window positions, and then splicing the time signals in the order of time windows from small to large to obtain the complete time signal of the source.

In the above operation process, the time signals in adjacent windows will partially overlap, and the length of the overlap is defined when the original observation signal is windowed and divided into frames at the beginning, which is half of the frame length. The processing of overlapping data in adjacent windows is generally to add the second half of the length of the previous window to the first half of the length of the latter window, and then divide by 2 to obtain the average.

S5. Perform wavelet decomposition on the complete time signal of the multi-channel independent source in the time domain to obtain wavelet coefficients at all levels;

Specifically, the formula for wavelet decomposition is as follows:

[c,1]=wavedec((Y,N,wname),

Among them, c is the wavelet decomposition vector, 1 represents the length of each level from high to low, Y represents the variable to be decomposed, N represents the number of layers of decomposition, and wname represents the generating function. In this embodiment, the number of layers for wavelet decomposition of multi-channel EOG data is ten, and the generating function of wavelet decomposition is db4.

S6, compare and analyze the wavelet coefficients at all levels and the evaluation criteria of slow eye movement, and identify the slow eye movement if they are consistent with the slow eye movement characteristics.

Specifically, the characteristics of slow eye movement are: (1) a slow sinusoidal offset appears in the signal for more than 1 second, that is, the signal frequency is lower than 1 Hz; (2) the initial movement speed of the signal is close to zero, in this embodiment The initial motion speed of the signal is less than 0.000001, which can be considered as close to zero; (3) EOG waveform does not have artifacts such as blinking, EEG and EMG. When the eye movement signal satisfies the above three conditions at the same time, it is considered that slow eye movement occurs in the eye movement signal. The comparison and judgment process in this embodiment will be described with reference to Fig. 9. Fig. 9B-(a) is a segment of slow eye movement extracted from the time domain signal:

The first row of Figure 9B-(a) is a waveform cut from the EOG signal returned to the time domain after blind separation in the frequency domain. The next two to six rows are the wavelet coefficients obtained after wavelet decomposition, respectively. It is the sixth layer wavelet, the seventh layer wavelet and so on until the tenth layer wavelet (marked in the figure). It can be seen from the figure that when Figure 9B-(a) is decomposed into the tenth layer of wavelet coefficients, the frequency of the signal is already lower than 1 Hz, the initial velocity is 0, and there is no artifact signal in the EOG waveform, so it can be judged that It is slow eye movement. On the other hand, D6 and D7 in Fig. 9B-(a) can easily see that the signal frequency is not lower than 1 Hz and there is a small amount of artifact signals, so they are not slow eye movements. The initial speed of D8 is not 0, and the signal frequency of D9 is higher than 1Hz, so it is not slow eye movement. To sum up, in the determination of slow eye movement, it is necessary to compare and analyze the wavelet coefficients at all levels with the evaluation criteria of slow eye movement, and only waveforms that meet all the characteristics at the same time can be identified as slow eye movement.

Since the time domain signals after inverse Fourier transform do not interfere with each other, wavelet analysis is performed on the multi-channel EOG signal in the time domain, and slow eye movement analysis is performed on the decomposed wavelet coefficients at all levels. Signal interference, high accuracy, less calculation, and can quickly extract slow eye movements from the EOG signal.

It should be noted that, as shown in Figure 5, the time-frequency domain waveforms of the six adjacent frequency points of the two independent sources obtained after the blind source separation of the two-channel EOG signal by the complex-valued ICA algorithm. The abscissa represents the position of the sliding window, and the ordinate represents the amplitude of the signal. It can be seen from the two waveform diagrams 5-(a) and 5-(b) that the third channel and the fifth channel in the top-to-bottom sequence have the problem of order blurring.

Figure 6 is the EOG signal waveforms obtained before and after the multi-channel eye movement data is separated by convolution ICA. The abscissa represents the sampling point, and the ordinate represents the amplitude of the signal. Comparing the two Figures 6-(a) and 6-(b), it can be seen that after the convolution ICA separation, the blinking artifact source signal is separated.

As shown in Figure 7, Figures 7(a) and 7(b) respectively represent the saccade EOG waveforms separated by linear ICA and convolutional ICA, where the abscissa shows the sampling point, and the ordinate is the amplitude of the signal . Figures 7(c) and 7(d) show the time-domain and frequency-domain waveform diagrams of the sweeping EOG signal of the second channel of Figures 7(a) and 7(b), respectively. The abscissa of the time-domain waveform diagram represents the sampling point, the ordinate represents the amplitude of the signal, the abscissa of the frequency-domain waveform diagram represents the frequency, and the ordinate represents the amplitude of the signal. It can be clearly seen from the two figures that after the linear ICA separation, the artifact signals are not separated "cleanly", and there are still blinking signals, and the saccade EOG signal separated by the linear ICA is higher than the saccade EOG signal separated by the convolution ICA. frequency bandwidth. Therefore, in this embodiment, the convolution ICA algorithm is preferably used to perform blind source separation processing on the eye movement data.

Figure 8 shows a schematic diagram of the average recognition rate of slow eye movement signals for different algorithms. The horizontal axis represents the order of the subjects, and the vertical axis represents the average recognition rate. It can be seen that the average recognition rate obtained by the convolution ICA method is 97.254%, which is 4.854%, 7.168% and 2.64% higher than the bandpass filtering method, the wavelet denoising method and the linear ICA method, respectively.

As shown in FIG. 9, FIG. 9A(a) and FIG. 9A(b) are time-domain EOG waveform diagrams after blind separation of convolution ICA and linear ICA, respectively. Wavelet decomposition is performed on each channel respectively, and it can be found that there are two waveforms in the fourth channel with slow eye movements (indicated by arrows), and the results are shown in Figures 9B(a) and 9B(b). It can be seen from the two figures in Fig. 9B that when decomposed to the tenth layer, there is slow eye movement. In order to compare the separation results of linear ICA, wavelet decomposition and analysis were performed on the waveforms of each channel after linear ICA separation where the waveform appeared slow eye movement after convolution separation. The experimental results are shown in Figures 9C and 9D. As can be seen from the four figures, no slow eye movements were present in the waveforms after linear ICA separation.

In addition, as shown in FIG. 10 , this embodiment also discloses a slow eye movement recognition system based on a convolutional hybrid model, including: a blind source separation module 10 , a scale compensation module 20 , a sorting module 30 , and a recovery module connected in sequence. module 40, wavelet decomposition module 50 and slow eye movement recognition module 60;

The blind source separation module 10 is used to perform blind source separation on the eye movement data of each frequency point in the frequency domain by using the complex-valued ICA algorithm, and obtain the frequency domain independent components of each independent source signal at the corresponding frequency point and separate the frequency domain. The components are transmitted to the scale compensation module 20;

The scale compensation module 20 is used to perform scale compensation on the independent components on each frequency point, restore the true proportional components of the independent components in the observed components, and transmit the compensated independent components to the sorting module 30;

The sorting module 30 is configured to use the constrained DOA algorithm to sort the compensated independent components, so that the independent sources on each frequency point are arranged according to the direction angle from small to large;

The recovery module 40 is used to perform inverse short-time Fourier transform processing on the independent components of each frequency point after scale compensation and sorting, to obtain the complete time signal of the multi-channel independent source in the time domain, and to obtain the complete time signal of the multi-channel independent source. The signal is transmitted to the wavelet decomposition module 50;

The wavelet decomposition module 50 is used to perform wavelet decomposition on the complete time signal of the multi-channel independent source in the time domain, obtain wavelet coefficients at all levels, and transmit the decomposition result to the slow eye movement recognition module 60;

The slow eye movement recognition module 60 is used to compare and analyze the wavelet coefficients at all levels with the slow eye movement evaluation criteria, and identify the slow eye movement as the characteristics of the slow eye movement.

It should be noted that the slow eye movement recognition system based on the convolution mixture model disclosed in this embodiment has the same or corresponding technical features and technical effects as the method disclosed in the foregoing embodiment, and details are not repeated here.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection of the present invention. within the range.

Claims (5)

1. a slow eye movement recognition method based on convolution hybrid model, is characterized in that, comprises: S1. In the frequency domain, the complex-valued ICA algorithm is used to perform blind source separation on the eye movement data of each frequency point to obtain the frequency domain independent components of each independent source signal at the corresponding frequency point; S2. Perform scale compensation on the independent components at each frequency point to restore the true proportional components of the independent components in the observed components; According to the separation matrix of each frequency point in the complex-valued ICA algorithm, the mixing matrix of the corresponding frequency point is obtained, wherein the separation matrix and the mixing matrix are mutually inverse matrices; Use the coefficients of the mixing matrix to compensate the independent components of each frequency point, and obtain the independent components of each frequency point after compensation; S3. The constrained DOA algorithm is used to sort and adjust the independent components after compensation, so that the independent sources on each frequency point are arranged according to the direction angle from small to large; a. Initialize an angle for each independent source; b. Calculate the different rows of each frequency point through the Root-Music algorithm to obtain the estimation of each source direction, where the rows of the separation matrix correspond to different independent sources; c. Set the proximity measure between the direction angle of each independent source and the initialization angle as ε(y, θ), and in the iterative process, determine whether the angle of each independent source is the same as the initialization angle; d. If they are the same, execute step e, and if they are not the same, execute step f; e. Set ε(y _j , θ _j ) to 0, and set the direction angle matrix T to calculate the adjustment matrix Q; f. Set ε(y _j , θ _j ) to 1, and return to the iterative process to recalculate the separation matrix W; S4, performing short-time inverse Fourier transform processing on the independent components of each frequency point after scale compensation and sorting, to obtain a complete time signal of multi-channel independent sources in the time domain; S5. Perform wavelet decomposition on the multi-channel independent sources in the time domain to obtain wavelet coefficients at all levels; S6, compare and analyze the wavelet coefficients at all levels and the evaluation criteria of slow eye movement, and identify the slow eye movement if they are consistent with the slow eye movement characteristics. 2 . The method according to claim 1 , wherein, in the step S5 , the generating function of the wavelet decomposition used is db4 , and the number of layers of the wavelet decomposition is ten. 3 . 3. The method according to claim 1, wherein the characteristics of the slow eye movement include: the frequency of the eye movement signal is lower than 1 Hz, the initial movement speed of the eye movement signal is approximately 0, and no false alarm occurs in the EOG signal. trace signal. 4. The method according to claim 1, characterized in that, before said step S1, further comprising: Collect multi-channel EOG data to obtain eye movement data in the time domain; Perform band-pass filtering and de-average processing on the eye movement data in the time domain to obtain the processed eye movement data; Perform short-time Fourier transform on the processed eye movement data, transform it from the time domain to the frequency domain, and obtain the eye movement data in the frequency domain. 5. A recognition system based on the slow eye movement recognition method based on the convolution hybrid model of one of claims 1-4, is characterized in that, comprises: the blind source separation module, the scale compensation module, the sorting module, Restoration module, wavelet decomposition module and slow eye movement recognition module; The blind source separation module is used to perform blind source separation on the eye movement data of each frequency point using the complex-valued ICA algorithm in the frequency domain, and obtain the frequency domain independent components of each independent source signal at the corresponding frequency point and separate the frequency domain independent components. Transfer to the scale compensation module; The scale compensation module is used to perform scale compensation for the independent components on each frequency point, restore the true proportional components of the independent components in the observed components, and transmit the compensated independent components to the sorting module; The sorting module is used to sort and adjust the compensated independent components by using the constrained DOA algorithm, so that the independent sources on each frequency point are arranged according to the direction angle from small to large; The recovery module is used to perform inverse short-time Fourier transform processing on the independent components of each frequency point after scale compensation and sorting, and obtain the complete time signal of the multi-channel independent source in the time domain and the complete time signal of the multi-channel independent source. It is transmitted to the wavelet decomposition module; The wavelet decomposition module is used to perform wavelet decomposition on the complete time signal of multi-channel independent sources in the time domain, obtain wavelet coefficients at all levels, and transmit the decomposition results to the slow eye movement recognition module; The slow eye movement recognition module is used to compare and analyze the wavelet coefficients at all levels with the slow eye movement evaluation criteria, and identify the slow eye movement if they are consistent with the slow eye movement characteristics.

Images