Abnormal EEG signal monitoring method based on artificial intelligence

Abstract

The present application provides an abnormal EEG signal monitoring method based on artificial intelligence, including: obtaining the EEG signal of the user; performing a scattering transformation on the EEG signal to obtain a set of scattering coefficients corresponding to the EEG signal; using a distributed fuzzy entropy measurement method to extract features from a one-dimensional discrete sequence determined based on the set of scattering coefficients to obtain signal features corresponding to the EEG signal; inputting the signal features into a preset EEG signal abnormality monitoring model to obtain monitoring results, wherein the EEG signal abnormality monitoring model is an improved extreme learning machine model constructed based on a single hidden layer feedforward neural network. The use of fuzzy distributed entropy technology to extract EEG signal features reduces the dependence on preset parameters such as embedding dimension and similarity tolerance. By improving the extreme learning machine, the time complexity of the algorithm is greatly reduced, and it has better robustness and stable generalization ability.

Description

Technical Field

[0001] The present application relates to the technical field of electroencephalogram (EEG) signal monitoring, and in particular to an artificial intelligence-based method for monitoring abnormal EEG signals. Background Art

[0002] At present, there is an increasing number of studies on EEG signal analysis and machine learning technology at home and abroad. In order to realize the classification of EEG signals and identify abnormal EEG signals in EEG signals, there is great potential for development in the field of computer-assisted medicine.

[0003] Artificial intelligence is a technology that uses the principles of statistics and computer science to improve algorithm performance by interpreting multivariate data rather than being limited to explicit instructions. It can be applied to the classification and recognition of EEG signals. However, existing EEG signal monitoring technology is too dependent on parameter design and sample richness, has high time complexity and poor robustness, which is not conducive to abnormal EEG signal monitoring. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide an EEG signal abnormality monitoring method based on artificial intelligence to reduce time complexity, improve robustness, reduce dependence on samples and parameter design, and facilitate abnormal monitoring of EEG signals.

[0005] In order to achieve the above purpose, the embodiments of the present application are implemented in the following manner:

[0006] In the first aspect, an embodiment of the present application provides an EEG signal abnormality monitoring method based on artificial intelligence, comprising: obtaining an EEG signal of a user; performing a scattering transform on the EEG signal to obtain a set of scattering coefficients corresponding to the EEG signal; using a distributed fuzzy entropy measurement method to perform feature extraction on a one-dimensional discrete sequence determined based on the scattering coefficient set to obtain a signal feature corresponding to the EEG signal; inputting the signal feature into a preset EEG signal abnormality monitoring model to obtain a monitoring result, wherein the EEG signal abnormality monitoring model is an improved extreme learning machine model constructed based on a single hidden layer feedforward neural network.

[0007] In combination with the first aspect, in a first possible implementation manner of the first aspect, performing a scattering transform on the EEG signal to obtain a scattering coefficient set corresponding to the EEG signal includes:

[0008] The EEG signal is subjected to scattering transformation using the following formula:

[0009]

[0010] Wherein, s is the EEG signal, and is a one-dimensional signal; for a icomplex wavelet functions of different scales, φ A is a low-pass filter, A is the scale; after the EEG signal is scattering transformed, there are 2 A scattering paths, each scattering path gets The scattering coefficient is Thus, a set of scattering coefficients corresponding to the EEG signal is obtained, wherein N is the number of sampling points of the EEG signal and M is the decomposition order.

[0011] In combination with the first aspect, in a second possible implementation method of the first aspect, a distributed fuzzy entropy measurement method is used to extract features from a one-dimensional discrete sequence determined based on the scattering coefficient set to obtain signal features corresponding to the EEG signal, including: determining a one-dimensional discrete sequence U={u(1),u(2),…,u(N)} based on the scattering coefficient set; reconstructing Nm m-dimensional vectors based on the one-dimensional discrete sequence U, where m is the embedding dimension; based on the Nm m-dimensional reconstructed vectors, determining the distance between any two reconstructed vectors and constructing a distance matrix; based on each element in the distance matrix, calculating a similarity matrix corresponding to the distance matrix; performing probability density estimation on the similarity matrix to obtain a probability density function; based on the probability density function, determining the fuzzy distribution entropy corresponding to the one-dimensional discrete sequence U as the signal feature corresponding to the EEG signal.

[0012] In combination with the second possible implementation manner of the first aspect, in a third possible implementation manner of the first aspect, reconstructing Nm m-dimensional vectors based on the one-dimensional discrete sequence U includes: reconstructing the j-th m-dimensional vector based on the one-dimensional discrete sequence U using the following formula:

[0013] V(j)={u(j),u(j+1),…,u(j+m-1)}-u0(i),

[0014] Where V(j) is the jth reconstruction vector, j = 1, 2, ..., N-m+1, u(j), u(j+1), ..., u(j+m-1) are the jth, j+1th, ..., j+m-1th elements in the one-dimensional discrete sequence U, respectively, and u0 is the reference mean of the reconstruction vector, that is:

[0015]

[0016] Among them, k=1,2,…,m-1, m>1.

[0017] In combination with the third possible implementation manner of the first aspect, in a fourth possible implementation manner of the first aspect, based on Nm m-dimensional reconstruction vectors, determining the distance between any two reconstruction vectors, and constructing a distance matrix, including: for any two reconstruction vectors, calculating the distance between the two reconstruction vectors V(p) and V(q):

[0018]

[0019] Among them, d pq is the distance between the reconstructed vectors V(p) and V(q), u(p+l) is the l+1th element in the reconstructed vector V(p), u0(p) is the mean of the reconstructed vector V(p), u(q+l) is the l+1th element in the reconstructed vector V(q), and u0(q) is the mean of the reconstructed vector V(q); based on the distance between each reconstructed vector and other reconstructed vectors, a distance matrix is ​​constructed:

[0020] D={d pq},

[0021] Among them, p,q=1,2,...,Nm.

[0022] In combination with the fourth possible implementation manner of the first aspect, in a fifth possible implementation manner of the first aspect, calculating a similarity matrix corresponding to the distance matrix based on each element in the distance matrix includes: calculating the fuzzy membership of each element in the distance matrix using the following formula:

[0023]

[0024] Among them, s pq Represents the element d in the distance matrix pq , r is the similarity tolerance of fuzzy distributed entropy, α>0;

[0025] Based on each element d of the distance matrix pq The corresponding fuzzy membership s pq , construct the similarity matrix:

[0026] S={s pq},

[0027] Among them, p,q=1,2,…,N-m+1.

[0028] In combination with the fifth possible implementation of the first aspect, in a sixth possible implementation of the first aspect, a probability density estimation is performed on the similarity matrix to obtain a probability density function, including: removing the diagonal elements of p=q in the similarity matrix to obtain a one-dimensional sequence S'={s'1, s'2, ..., s' t ,…,s′T}, T = (N-m + 1) (Nm); Based on the one-dimensional sequence S', the following formula is used to estimate the probability density and obtain the probability density function:

[0029]

[0030] in, Represents the probability density function of the population, s is the independent variable, and h is the hyperparameter.

[0031] In combination with the sixth possible implementation manner of the first aspect, in a seventh possible implementation manner of the first aspect, based on the probability density function, determining the fuzzy distribution entropy corresponding to the one-dimensional discrete sequence U includes: based on the probability density function, determining the fuzzy distribution entropy of the one-dimensional discrete sequence U as:

[0032]

[0033] in, is the fuzzy distribution entropy.

[0034] In combination with the first aspect, in an eighth possible implementation manner of the first aspect, a process of constructing an improved extreme learning machine model based on a single hidden layer feedforward neural network is as follows:

[0035] Constructing a single hidden layer feedforward neural network, wherein the single hidden layer feedforward neural network includes an input layer, a hidden layer and an output layer;

[0036] Establish the objective function:

[0037]

[0038] Among them, L is the number of hidden layer nodes, β is the output weight of the hidden layer, and C is the penalty factor. is the number of input samples, ξ i is the i-th sample x i The error between the hidden layer output and h(x i ) is the sample x i The corresponding hidden layer output vector, For sample x i The corresponding output layer ideal output r i The transposed vector of ;

[0039] According to the KKT condition, the Lagrangian function is constructed to solve the objective function, which is equivalent to:

[0040]

[0041] Where Y is the dimension of the output vector, ρ i,j is the Lagrange multiplier, β j is the j-th dimension element of the output weight, ri,j For sample x i The ideal output r i The j-th element of i,j For sample x i The error ξ i The j-th dimension element of ;

[0042] Solve the least squares solution after optimizing the constraints:

[0043]

[0044] in, is the updated output weight, H is the hidden layer output matrix, H T is the transpose of the hidden layer output matrix, I is the identity matrix, and R is the output layer output matrix;

[0045] Complete the construction of the improved extreme learning machine model.

[0046] In combination with the eighth possible implementation manner of the first aspect, in a ninth possible implementation manner of the first aspect, the method further includes:

[0047] The penalty factor C is set as an adaptive penalty factor, and the hidden layer output matrix solution formula of the improved extreme learning machine model is:

[0048]

[0049] Among them, Δ(x(r), r) is the nonlinear adaptive adjustment parameter, K is the kernel function, and y is -dimensional compensation vector, containing the elements Ω ELM =HH T is the kernel matrix;

[0050]

[0051] in, is a given constant and θ(r) is a nonlinear function.

[0052] In the second aspect, an embodiment of the present application provides an EEG signal abnormality monitoring device based on artificial intelligence, comprising: a signal acquisition unit, used to acquire the user's EEG signal; a scattering transformation unit, used to perform a scattering transformation on the EEG signal to obtain a scattering coefficient set corresponding to the EEG signal; a feature extraction unit, used to perform feature extraction on a one-dimensional discrete sequence determined based on the scattering coefficient set using a distributed fuzzy entropy measurement method to obtain a signal feature corresponding to the EEG signal; a signal monitoring unit, used to input the signal feature into a preset EEG signal abnormality monitoring model to obtain a monitoring result, wherein the EEG signal abnormality monitoring model is an improved extreme learning machine model constructed based on a single hidden layer feedforward neural network.

[0053] Beneficial effects:

[0054] By acquiring the user's EEG signal, performing scattering transformation to obtain a scattering coefficient set, and then using the distributed fuzzy entropy measurement method to extract features from the one-dimensional discrete sequence determined based on the scattering coefficient set, the signal features are obtained, and input into the preset EEG signal abnormality monitoring model (an improved extreme learning machine model based on a single hidden layer feedforward neural network) to obtain the monitoring results. 1. This method uses fuzzy distributed entropy technology to extract EEG signal features, which reduces the dependence on preset parameters such as embedding dimension and similarity tolerance. By improving the extreme learning machine, the time complexity of the algorithm is greatly reduced, and it has better robustness and stable generalization ability. 2. Considering the element distance of each reconstructed vector, a distance matrix is ​​constructed based on this, which can well reflect the reconstruction characteristics of the EEG signal. 3. In the process of calculating the similarity matrix, the membership of each element can be measured more accurately, and the similarity matrix is ​​constructed based on this, which can effectively measure the characteristics of the EEG signal, and further use the non-parametric density estimation method to perform probability density estimation, and can eliminate the deviation caused by the self-matching of the reconstructed vector. 4. An improved extreme learning machine model is constructed using a single hidden layer feedforward neural network, and the mapping of the extreme learning machine is realized using kernel function mapping. This greatly improves the problems of poor stability and poor algorithm generalization ability caused by random selection of hidden layer parameters, greatly reduces the time complexity of the algorithm, and can quickly obtain the least squares optimal solution without optimizing the number of hidden layer nodes. It uses nonlinear adaptive adjustment parameters to replace the setting of the penalty factor to avoid problems caused by overfitting, thereby achieving the optimal classification effect without relying on parameter setting experience.

[0055] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 A schematic diagram of the structure of the EEG signal abnormality monitoring model provided in an embodiment of the present application.

[0058] Figure 2 A flowchart of an artificial intelligence-based EEG signal abnormality monitoring method provided in an embodiment of the present application.

[0059] Figure 3 This is a structural block diagram of an artificial intelligence-based EEG signal abnormality monitoring device provided in an embodiment of the present application.

[0060] Icons: 10- artificial intelligence-based EEG signal abnormality monitoring device; 11- signal acquisition unit; 12- scattering transformation unit; 13- feature extraction unit; 14- signal monitoring unit. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0062] In order to facilitate the understanding of this solution, the construction process of the EEG signal abnormality monitoring model involved in this solution is first introduced here.

[0063] See also Figure 1 , Figure 1 A schematic diagram of the structure of the EEG signal abnormality monitoring model provided in an embodiment of the present application.

[0064] The extreme learning machine with a single hidden layer has more advantages in learning rate and generalization ability. Therefore, this scheme is based on a single hidden layer feedforward neural network and uses the improved extreme learning machine algorithm to realize the construction of an abnormal EEG signal monitoring model.

[0065] In this embodiment, the process of constructing an improved extreme learning machine model based on a single hidden layer feedforward neural network is as follows:

[0066] Firstly, a single hidden layer feedforward neural network is constructed, wherein the single hidden layer feedforward neural network includes an input layer, a hidden layer and an output layer.

[0067] The output equation of the jth hidden layer node is defined as:

[0068] h j (x) = G(a j , bj , x), (1)

[0069] Among them, h j (x) is the output of the jth hidden layer node, G is the activation function, a j is the input weight corresponding to the jth hidden layer node, b j is the bias corresponding to the jth hidden layer node, and x is the input of the single hidden layer feedforward neural network. Here, h j The function of (x) is to map the data of the input layer from its original space to the feature space of the extreme learning machine.

[0070] Based on this, the hidden layer output with L hidden layer nodes is:

[0071] h(x)=[G(h1(x),…,h L (x))], (2)

[0072] By input Taking training samples as an example, the hidden layer output matrix of the extreme learning machine is:

[0073]

[0074] Among them, H is the hidden layer output matrix of the extreme learning machine, L is the number of hidden layer nodes, is the number of input samples, Respectively samples (input training samples).

[0075] Therefore, the output expression of the extreme learning machine with L hidden layer nodes is:

[0076]

[0077] Among them, f L is the output of the extreme learning machine with L hidden layer nodes, β j It is the j-th dimensional element of the output weight, that is, the output weight β of the hidden layer (a matrix of L×Y, where L is the number of hidden layer nodes and Y is the dimension of the output vector) corresponding to the output weight of the j-th hidden layer node.

[0078] Let the overall output matrix of the extreme learning machine be:

[0079] R=Hβ, (5)

[0080] Among them, R is the overall output matrix of the extreme learning machine.

[0081] Then the overall output matrix R can be expressed as:

[0082]

[0083] The overall output matrix R is the training target. Correspondingly, we can establish

[0084]

[0085] Among them, L is the number of hidden layer nodes, β is the output weight of the hidden layer, and C is the penalty factor. is the number of input samples, ξ i is the i-th sample x i The error between the hidden layer output and h(x i ) is the sample x i The corresponding hidden layer output vector, For sample x i The corresponding output layer ideal output r i The transposed vector of .

[0086] According to the KKT (Karush-Kuhn-Tucker) condition, the Lagrangian function is constructed to solve the objective function, which is equivalent to:

[0087]

[0088] Where Y is the dimension of the output vector, ρ i,j is the Lagrange multiplier, r i,j For sample x i The corresponding ideal output r i The j-th element of i,j For sample x i The corresponding error ξ i The j-th element of .

[0089] After optimizing the constraints, we can get:

[0090]

[0091] Among them, I is the identity matrix, C is the penalty factor, H is the hidden layer output matrix, and H T is the transpose of the hidden layer output matrix, and ρ is the Lagrange multiplier ρ i,j The matrix composed of , R is the output matrix of the output layer.

[0092] Based on this, the least squares solution can be solved:

[0093]

[0094] in, is the updated output weight, H is the hidden layer output matrix, H T is the transpose of the hidden layer output matrix, I is the identity matrix, and R is the output layer output matrix.

[0095] When the penalty factor C is too large, it will be infinitely close to the traditional extreme learning machine without penalty factor. By setting a suitable penalty factor C, the problem of overfitting can be avoided, thereby achieving the best classification effect. However, the value of the penalty factor C is heavily dependent on the empirical setting. It is difficult to determine a suitable penalty factor by adjustment in the extreme learning machine. Therefore, an adaptive penalty factor is designed here to adjust the output weight of the extreme learning machine online.

[0096] First, the penalty factor C is set to an adaptive penalty factor, and the solution formula for the hidden layer output matrix H of the improved extreme learning machine model is:

[0097]

[0098] Among them, f(x) is the output expression of the extreme learning machine, Δ(x(r), r) is the nonlinear adaptive adjustment parameter, and h(x) is the hidden layer output.

[0099] make:

[0100] h(x i )=[G(w i1 ·x i +b i1 ),…,G(w iL ·x i +b iL )], (12)

[0101] Among them, w iL For sample x i The corresponding input weight of the Lth hidden layer node, b L For sample x i The corresponding bias of the Lth hidden layer node.

[0102] The hidden layer output matrix H of the improved extreme learning machine model can be written as follows:

[0103]

[0104] Writing the hidden layer output matrix H as LMI (linear matrix inequality) can be expressed as:

[0105] H i =Ax i +B, (14)

[0106] Among them, A and B are both constant matrices.

[0107] All hidden layer outputs h(x i ) is considered as sample x i The nonlinear mapping of is:

[0108]

[0109] Where E is the mapping of the constant matrix B, and I is the identity matrix. Then, it can be seen that equation (14) is h(x i Based on this, in this embodiment, a kernel mapping is constructed to replace HH T ,have:

[0110] HH T (i, j) = (K(x i , x j )+y) z , (16)

[0111] Among them, K is the kernel function, y is the compensation vector, and z is the order.

[0112] Then, the hidden layer output matrix solution formula of the improved extreme learning machine model can be converted to:

[0113]

[0114] Among them, Ω ELM =HH T is the kernel matrix, is a given constant and θ(r) is a nonlinear function.

[0115] By inputting training samples (signal features corresponding to EEG signals), the improved extreme learning machine model can be trained, thereby completing the construction of the improved extreme learning machine model (EEG signal abnormality monitoring model).

[0116] An improved extreme learning machine model is constructed using a single hidden layer feedforward neural network, and the mapping of the extreme learning machine is realized by kernel function mapping. This greatly improves the problems of poor stability and poor generalization ability of the algorithm caused by random selection of hidden layer parameters, greatly reduces the time complexity of the algorithm, and can quickly obtain the least squares optimal solution without optimizing the number of hidden layer nodes. Nonlinear adaptive adjustment parameters are used to replace the setting of penalty factors to avoid problems caused by overfitting, thereby achieving the optimal classification effect without relying on parameter setting experience.

[0117] After obtaining the EEG signal abnormality monitoring model, the EEG signal abnormality monitoring method based on artificial intelligence can be run accordingly to realize abnormal monitoring of EEG signals. The execution subject of the EEG signal abnormality monitoring method based on artificial intelligence can be a server or a monitoring terminal, which is collectively referred to as electronic equipment here without limitation.

[0118] See also Figure 2 , Figure 2Flow chart of the method for monitoring abnormal EEG signals based on artificial intelligence. In this embodiment, the method for monitoring abnormal EEG signals based on artificial intelligence may include step S10, step S20, step S30 and step S40.

[0119] First, the electronic device may execute step S10.

[0120] Step S10: Acquire the user's EEG signal.

[0121] In order to monitor abnormalities in the user's EEG signals, the electronic device may acquire the user's EEG signals. The user's EEG signals may be acquired in real time or in advance using an EEG signal sensing device, which is not limited here.

[0122] After acquiring the user's EEG signal, the electronic device may execute step S20.

[0123] Step S20: performing a scattering transform on the EEG signal to obtain a scattering coefficient set corresponding to the EEG signal.

[0124] In this embodiment, the electronic device can perform scattering transformation on the EEG signal.

[0125] Exemplarily, the electronic device may use the following formula to perform scattering transformation on the EEG signal:

[0126]

[0127] Among them, s is the EEG signal, and it is a one-dimensional signal; …、 for a i complex wavelet functions of different scales, φ A is a low-pass filter and A is the scale.

[0128] In this embodiment, the scattering transform operation mainly includes three parts: multi-scale complex wavelet transform, modular operation and local mean calculation. Multi-scale complex wavelet transform is mainly used to obtain the underlying characteristics of the signal, modular operation retains high-frequency information and ensures time-shift invariance by calculating the modularity of the high-frequency coefficients of the signal, and local mean calculation is used to obtain relatively stable frequency characteristics.

[0129] After the EEG signal is scatter transformed, there are 2 A scattering paths, each scattering path gets The scattering coefficient is Thus, the scattering coefficient set corresponding to the EEG signal is obtained, where N is the number of sampling points of the EEG signal, M is the decomposition order, and This is the permutation and combination calculation method.

[0130] After obtaining the scattering coefficient set corresponding to the EEG signal, the electronic device may execute step S30.

[0131] Step S30: using a distributed fuzzy entropy measurement method to perform feature extraction on a one-dimensional discrete sequence determined based on the scattering coefficient set, to obtain signal features corresponding to the EEG signal.

[0132] In this embodiment, the electronic device may determine a one-dimensional discrete sequence based on the scattering coefficient set:

[0133] U={u(1),u(2),...,u(N)}, (19)

[0134] Wherein, U is a one-dimensional discrete sequence, and u(1), u(2), ..., u(N) are each element in the one-dimensional discrete sequence U, that is, each scattering coefficient in the scattering coefficient set.

[0135] After determining the one-dimensional discrete sequence U, the electronic device can reconstruct Nm m-dimensional vectors based on the one-dimensional discrete sequence U, where m is the embedding dimension.

[0136] Exemplarily, the electronic device may reconstruct the j-th m-dimensional vector based on the one-dimensional discrete sequence U using the following formula:

[0137] V(j)={u(j),u(j+1),...,u(j+m-1)}-u0(i), (20)

[0138] Where V(j) is the jth reconstruction vector, j = 1, 2, ..., N-m+1, u(j), u(j+1), ..., u(j+m-1) are the jth, j+1th, ..., j+m-1th elements in the one-dimensional discrete sequence U, respectively, and u0 is the reference mean of the reconstruction vector, that is:

[0139]

[0140] Among them, k=1,2,…,m-1, m>1.

[0141] In this way, phase space reconstruction can be achieved, and different orders can be taken into consideration when obtaining the reference mean of the reconstructed vector, so that the reconstructed vector can well reflect the characteristics of the EEG signal and is suitable for abnormal monitoring of EEG signals of different abnormal types.

[0142] After reconstructing the phase space, the electronic device can further determine the distance between any two reconstruction vectors based on the Nm m-dimensional reconstruction vectors and construct a distance matrix.

[0143] Exemplarily, for any two reconstruction vectors, the electronic device may calculate the distance between the two reconstruction vectors V(p) and V(q):

[0144]

[0145] Among them, d pq is the distance between the reconstructed vectors V(p) and V(q), u(p+l) is the l+1th element in the reconstructed vector V(p), u0(p) is the mean of the reconstructed vector V(p), u(q+l) is the l+1th element in the reconstructed vector V(q), and u0(q) is the mean of the reconstructed vector V(q).

[0146] After calculating the distance between each reconstructed vector and other reconstructed vectors, the electronic device can construct a distance matrix based on the distance between each reconstructed vector and other reconstructed vectors:

[0147] D={d pq}, (twenty three)

[0148] Among them, p,q=1,2,…,Nm.

[0149] Considering the element distance of each reconstructed vector and constructing a distance matrix based on it, the reconstruction characteristics of the EEG signal can be well reflected.

[0150] After constructing the distance matrix, the electronic device can perform probability density estimation on the similarity matrix.

[0151] Exemplarily, the electronic device may use the following formula to calculate the fuzzy membership of each element in the distance matrix:

[0152]

[0153] Among them, s pq Represents the element d in the distance matrix pq , r′ is the similarity tolerance of fuzzy distributed entropy, α>0.

[0154] Then, the electronic device can be based on each element d in the distance matrix pq The corresponding fuzzy membership s pq , construct the similarity matrix:

[0155] S={s pq}, (25)

[0156] Among them, p,q=1,2,…,N-m+1.

[0157] After obtaining the similarity matrix, the electronic device can perform probability density estimation to obtain a probability density function.

[0158] Exemplarily, the electronic device may remove the diagonal elements of p=q in the similarity matrix to obtain a one-dimensional sequence S' composed of the remaining elements:

[0159] S'={s'1,s'2,…,s' t ,…,s′ T}0, (26)

[0160] Among them, T = (N-m+1)(Nm).

[0161] Then, based on the one-dimensional sequence S', the probability density can be estimated using the following formula to obtain the probability density function:

[0162]

[0163] in, Represents the probability density function of the population, s is the independent variable, and h is the hyperparameter.

[0164] In the process of similarity matrix calculation, the membership of each element can be measured more accurately. Based on this, the similarity matrix can be constructed, which can effectively measure the characteristics of EEG signals, and further use non-parametric density estimation to perform probability density estimation, and eliminate the deviation caused by self-matching of the reconstructed vector.

[0165] After determining the probability density function, the electronic device can determine the fuzzy distribution entropy corresponding to the one-dimensional discrete sequence U based on the probability density function as the signal feature corresponding to the EEG signal.

[0166] Exemplarily, the electronic device may determine the fuzzy distribution entropy of the one-dimensional discrete sequence U based on the probability density function as follows:

[0167]

[0168] in, is the fuzzy distributed entropy, m is the dimension, α>0, r′ is the similarity tolerance of the fuzzy distributed entropy, and T is the total number of elements in the one-dimensional sequence S′.

[0169] The fuzzy distribution entropy corresponding to the one-dimensional discrete sequence U is the signal feature corresponding to the EEG signal, which can be used as the input of the EEG signal abnormality monitoring model.

[0170] Based on this, the electronic device may execute step S40.

[0171] Step S40: inputting the signal feature into a preset EEG signal abnormality monitoring model to obtain a monitoring result, wherein the EEG signal abnormality monitoring model is an improved extreme learning machine model constructed based on a single hidden layer feedforward neural network.

[0172] In this embodiment, the electronic device can input the obtained signal characteristics (fuzzy distribution entropy) into a preset EEG signal abnormality monitoring model, identify the signal characteristics through the EEG signal abnormality monitoring model, output the monitoring results, and determine whether the EEG signal is abnormal, thereby realizing abnormal EEG signal monitoring of the user.

[0173] By acquiring the user's EEG signal, performing scattering transformation to obtain a scattering coefficient set, and then using the distributed fuzzy entropy measurement method to extract features from the one-dimensional discrete sequence determined based on the scattering coefficient set, the signal features are obtained and input into the preset EEG signal abnormality monitoring model (an improved extreme learning machine model based on a single hidden layer feedforward neural network) to obtain the monitoring results. This method uses fuzzy distributed entropy technology to extract EEG signal features, reducing the dependence on preset parameters such as embedding dimension and similarity tolerance. By improving the extreme learning machine, the time complexity of the algorithm is greatly reduced, and it has better robustness and stable generalization ability.

[0174] See also Figure 3 Based on the same inventive concept, the embodiment of the present application also provides an artificial intelligence-based EEG signal abnormality monitoring device 10, comprising:

[0175] The signal acquisition unit 11 is used to acquire the user's electroencephalogram signal.

[0176] The scattering transformation unit 12 is used to perform scattering transformation on the EEG signal to obtain a set of scattering coefficients corresponding to the EEG signal.

[0177] The feature extraction unit 13 is used to extract features from the one-dimensional discrete sequence determined based on the scattering coefficient set by using a distributed fuzzy entropy measurement method to obtain signal features corresponding to the EEG signal.

[0178] The signal monitoring unit 14 is used to input the signal characteristics into a preset EEG signal abnormality monitoring model to obtain a monitoring result, wherein the EEG signal abnormality monitoring model is an improved extreme learning machine model constructed based on a single hidden layer feedforward neural network.

[0179] In this embodiment, the scattering transformation unit 12 is specifically used to perform scattering transformation on the EEG signal using the following formula:

[0180]

[0181] Wherein, s is the EEG signal, and is a one-dimensional signal; ψ a1 ,…,ψ ai for ai complex wavelet functions of different scales, φ A is a low-pass filter, A is the scale; after the EEG signal is scattering transformed, there are 2A scattering paths, each scattering path gets The scattering coefficient is Thus, a set of scattering coefficients corresponding to the EEG signal is obtained, wherein N is the number of sampling points of the EEG signal and M is the decomposition order.

[0182] In this embodiment, the feature extraction unit 13 is specifically used to determine a one-dimensional discrete sequence U = {u(1), u(2), ..., u(N)} based on the scattering coefficient set; reconstruct Nm m-dimensional vectors based on the one-dimensional discrete sequence U, where m is the embedding dimension; determine the distance between any two reconstructed vectors based on the Nm m-dimensional reconstructed vectors, and construct a distance matrix; calculate the similarity matrix corresponding to the distance matrix based on each element in the distance matrix; perform probability density estimation on the similarity matrix to obtain a probability density function; based on the probability density function, determine the fuzzy distribution entropy corresponding to the one-dimensional discrete sequence U as the signal feature corresponding to the EEG signal.

[0183] In this embodiment, the feature extraction unit 13 is specifically used to reconstruct the j-th m-dimensional vector based on the one-dimensional discrete sequence U using the following formula:

[0184] V(j)={u(j),u(j+1),…,u(j+m-1)}-u0(i),

[0185] Where V(j) is the jth reconstruction vector, j = 1, 2, ..., N-m+1, u(j), u(j+1), ..., u(j+m-1) are the jth, j+1th, ..., j+m-1th elements in the one-dimensional discrete sequence U, respectively, and u0 is the reference mean of the reconstruction vector, that is:

[0186]

[0187] Among them, k=1,2,…,m-1, m>1.

[0188] In this embodiment, the feature extraction unit 13 is specifically used to calculate the distance between two reconstruction vectors V(p) and V(q) for any two reconstruction vectors:

[0189]

[0190] Among them, d pqis the distance between the reconstructed vectors V(p) and V(q), u(p+l) is the l+1th element in the reconstructed vector V(p), u0(p) is the mean of the reconstructed vector V(p), u(q+l) is the l+1th element in the reconstructed vector V(q), and u0(q) is the mean of the reconstructed vector V(q); based on the distance between each reconstructed vector and other reconstructed vectors, a distance matrix is ​​constructed:

[0191] D={d pq},

[0192] Among them, p,q=1,2,…,Nm.

[0193] In this embodiment, the feature extraction unit 13 is specifically used to calculate the fuzzy membership of each element in the distance matrix using the following formula:

[0194]

[0195] Among them, s pq Represents the element d in the distance matrix pq , r is the similarity tolerance of fuzzy distributed entropy, α>0; based on each element d in the distance matrix pq The corresponding fuzzy membership s pq , construct the similarity matrix:

[0196] S={s pq},

[0197] Among them, p,q=1,2,...,N-m+1.

[0198] In this embodiment, the feature extraction unit 13 is specifically used to remove the diagonal elements of p=q in the similarity matrix to obtain a one-dimensional sequence composed of the remaining elements:

[0199] S'={s'1,s'2,…,s' t ,...,s′ T}, T = (N-m + 1)(Nm),

[0200] Based on the one-dimensional sequence S', the following formula is used to estimate the probability density and obtain the probability density function:

[0201]

[0202] in, Represents the probability density function of the population, s is the independent variable, and h is the hyperparameter.

[0203] In this embodiment, the feature extraction unit 13 is specifically used to determine the fuzzy distribution entropy of the one-dimensional discrete sequence U based on the probability density function:

[0204]

[0205] in, is the fuzzy distribution entropy.

[0206] To summarize, the embodiment of the present application provides an EEG signal abnormality monitoring method based on artificial intelligence. The method obtains the user's EEG signal, performs a scattering transform to obtain a scattering coefficient set, and then uses a distributed fuzzy entropy measurement method to extract features of a one-dimensional discrete sequence determined based on the scattering coefficient set to obtain signal features. The signal features are input into a preset EEG signal abnormality monitoring model (an improved extreme learning machine model constructed based on a single hidden layer feedforward neural network) to obtain monitoring results.

[0207] 1. This method uses fuzzy distributed entropy technology to extract EEG signal features, reducing the dependence on preset parameters such as embedding dimension and similarity tolerance. By improving the extreme learning machine, the time complexity of the algorithm is greatly reduced, and it has better robustness and stable generalization ability.

[0208] 2. Considering the element distance of each reconstructed vector and constructing a distance matrix based on it, it can well reflect the reconstruction characteristics of the EEG signal.

[0209] 3. In the process of similarity matrix calculation, the membership of each element can be measured more accurately. Based on this, the similarity matrix is ​​constructed, which can effectively measure the characteristics of EEG signals, and further use non-parametric density estimation to perform probability density estimation, and eliminate the deviation caused by self-matching of the reconstructed vector.

[0210] 4. An improved extreme learning machine model is constructed using a single hidden layer feedforward neural network, and the mapping of the extreme learning machine is realized using kernel function mapping. This greatly improves the problems of poor stability and poor algorithm generalization ability caused by random selection of hidden layer parameters, greatly reduces the time complexity of the algorithm, and can quickly obtain the least squares optimal solution without optimizing the number of hidden layer nodes. It uses nonlinear adaptive adjustment parameters to replace the setting of the penalty factor to avoid problems caused by overfitting, thereby achieving the optimal classification effect without relying on parameter setting experience.

[0211] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0212] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for monitoring abnormal EEG signals based on artificial intelligence, characterized in that: include: Obtain the user's EEG signal; Performing a scattering transform on the EEG signal to obtain a set of scattering coefficients corresponding to the EEG signal; Using a distributed fuzzy entropy measurement method to extract features from a one-dimensional discrete sequence determined based on the scattering coefficient set, to obtain signal features corresponding to the EEG signal; Inputting the signal feature into a preset EEG signal abnormality monitoring model to obtain a monitoring result, wherein the EEG signal abnormality monitoring model is an improved extreme learning machine model constructed based on a single hidden layer feedforward neural network; The distributed fuzzy entropy measurement method is used to extract features from the one-dimensional discrete sequence determined based on the scattering coefficient set to obtain signal features corresponding to the EEG signal, including: Determine a one-dimensional discrete sequence U={u(1),u(2),…,u(N)} based on the scattering coefficient set; Reconstruct N-m+1 m-dimensional vectors based on the one-dimensional discrete sequence U, where m is the embedding dimension; Based on N-m+1 m-dimensional reconstruction vectors, determine the distance between any two reconstruction vectors and construct a distance matrix; Based on each element in the distance matrix, a similarity matrix corresponding to the distance matrix is ​​calculated; Performing probability density estimation on the similarity matrix to obtain a probability density function; Based on the probability density function, determining the fuzzy distribution entropy corresponding to the one-dimensional discrete sequence U as the signal feature corresponding to the EEG signal; Reconstruct N-m+1 m-dimensional vectors based on the one-dimensional discrete sequence U, including: Based on the one-dimensional discrete sequence U, the j-th m-dimensional vector is reconstructed using the following formula: V(j)={u(j),u(j+1),…,u(j+m-1)}-u0(i), Where V(j) is the jth reconstruction vector, j = 1, 2, ..., N-m+1, u(j), u(j+1), ..., u(j+m-1) are the jth, j+1th, ..., j+m-1th elements in the one-dimensional discrete sequence U, respectively, and u0 is the reference mean of the reconstruction vector, that is: Where k = 0, 1, 2, ..., m-1, m> 1; Based on N-m+1 m-dimensional reconstruction vectors, the distance between any two reconstruction vectors is determined and a distance matrix is ​​constructed, including: For any two reconstructed vectors, calculate the distance between the two reconstructed vectors V(p) and V(q): Among them, d pq is the distance between the reconstructed vectors V(p) and V(q), u(p+l) is the l+1th element in the reconstructed vector V(p), u0(p) is the mean of the reconstructed vector V(p), u(q+l) is the l+1th element in the reconstructed vector V(q), and u0(q) is the mean of the reconstructed vector V(q); Based on the distance between each reconstructed vector and other reconstructed vectors, a distance matrix is ​​constructed: D={d pq }, Where p,q=1,2,…,N-m+1; Based on each element in the distance matrix, a similarity matrix corresponding to the distance matrix is ​​calculated, including: The fuzzy membership of each element in the distance matrix is ​​calculated using the following formula: Among them, s pq Represents the element d in the distance matrix pq , r is the similarity tolerance of fuzzy distributed entropy, α>0; Based on each element d of the distance matrix pq The corresponding fuzzy membership s pq , construct the similarity matrix: S={s pq }, Among them, p,q=1,2,…,N-m+1.

2. The method for abnormal EEG signal monitoring based on artificial intelligence according to claim 1, characterized in that: Performing a scattering transform on the EEG signal to obtain a scattering coefficient set corresponding to the EEG signal includes: The EEG signal is subjected to scattering transformation using the following formula: Wherein, s is the EEG signal, and is a one-dimensional signal; for a i complex wavelet functions of different scales, φ A is a low-pass filter, A is the scale; After the EEG signal is scatter transformed, there are 2 A scattering paths, each scattering path gets The scattering coefficient is Thus, a set of scattering coefficients corresponding to the EEG signal is obtained, wherein N is the number of sampling points of the EEG signal and M is the decomposition order.

3. The method for monitoring abnormal EEG signals based on artificial intelligence according to claim 1, characterized in that: Probability density estimation is performed on the similarity matrix to obtain a probability density function, including: Eliminate the diagonal elements of p=q in the similarity matrix to obtain a one-dimensional sequence S'={s'1,s'2,...,s' t ,…,s′ T }, T = (N-m+1)(Nm); Based on the one-dimensional sequence S', the following formula is used to estimate the probability density and obtain the probability density function: in, Represents the probability density function of the population, s is the independent variable, and h is the hyperparameter.

4. The method for monitoring abnormal EEG signals based on artificial intelligence according to claim 3, characterized in that: Based on the probability density function, the fuzzy distribution entropy corresponding to the one-dimensional discrete sequence U is determined, including: Based on the probability density function, the fuzzy distribution entropy of the one-dimensional discrete sequence U is determined as: in, is the fuzzy distribution entropy.

5. The method for abnormal EEG signal monitoring based on artificial intelligence according to claim 1, characterized in that: The process of constructing an improved extreme learning machine model based on a single hidden layer feedforward neural network is as follows: Constructing a single hidden layer feedforward neural network, wherein the single hidden layer feedforward neural network includes an input layer, a hidden layer and an output layer; Establish the objective function: Among them, L is the number of hidden layer nodes, β is the output weight of the hidden layer, and C is the penalty factor. is the number of input samples, ξ i is the i-th sample x i The error between the hidden layer output and h(x i ) is the sample x i The corresponding hidden layer output vector, For sample x i The corresponding output layer ideal output r i The transposed vector of ; According to the KKT condition, the Lagrangian function is constructed to solve the objective function, which is equivalent to: Where Y is the dimension of the output vector, ρ i,j is the Lagrange multiplier, β j is the j-th dimension element of the output weight, r i,j For sample x i The ideal output r i The j-th element of i,j For sample x i The error ξ i The j-th dimension element of ; Solve the least squares solution after optimizing the constraints: in, is the updated output weight, H is the hidden layer output matrix, H T is the transpose of the hidden layer output matrix, I is the identity matrix, and R is the output layer output matrix; Complete the construction of the improved extreme learning machine model.

6. The method for monitoring abnormal EEG signals based on artificial intelligence according to claim 5, characterized in that: The method further comprises: The penalty factor C is set as an adaptive penalty factor, and the hidden layer output matrix solution formula of the improved extreme learning machine model is: Among them, Δ(x(r), r) is the nonlinear adaptive adjustment parameter, K is the kernel function, and y is -dimensional compensation vector, containing the elements Ω ELM =HH T is the kernel matrix; in, is a given constant and θ(r) is a nonlinear function.

Images