Real-time electroencephalogram signal denoising method based on deep neural network feedback control

Abstract

The invention discloses an electroencephalogram signal real-time denoising method based on deep neural network feedback control. The electroencephalogram signal real-time denoising method comprises the steps that a noisy electroencephalogram signal and a noise reference signal are collected through a composite electrode; performing high-pass filtering processing on the noise reference signal, filtering out low-frequency artifacts, and retaining high-frequency myoelectricity noise components; inputting the filtered noise reference signal into a deep neural network through a tapped delay line, and synchronously inputting the noise-containing electroencephalogram signal into the deep neural network; the deep neural filter performs modeling on the noise-containing electroencephalogram signal by using a deep neural network to generate a removal signal opposite to the noise signal, and subtracts the noise-containing electroencephalogram signal from the removal signal to obtain a denoised electroencephalogram signal; applying the de-noised electroencephalogram signals to learning of a neural network, wherein the learning is completed through error back propagation; the method not only has an efficient noise suppression effect, but also has high environmental adaptability and calculation efficiency, and has higher real-time adaptability and better denoising quality.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal denoising, and particularly to a real-time EEG signal denoising method based on deep neural network feedback control. Background Technique

[0002] As a non-invasive means of monitoring brain activity, EEG signals have been widely used in fields such as neuroscience, clinical medicine, and brain-computer interfaces. However, during actual acquisition, EEG signals are easily interfered by various noises, such as electromagnetic noise, electromyographic noise, and eye movement artifacts. These noises will seriously affect the quality of EEG signals, reduce the signal-to-noise ratio of the signals, and thus affect the accurate interpretation and subsequent analysis of the signals. Therefore, how to effectively remove the noise in EEG signals has become an urgent problem to be solved.

[0003] Traditional denoising methods, such as low-pass filters and band-pass filters, can filter out low-frequency and high-frequency noises to a certain extent, but have poor effects on electromyographic noise or other complex non-stationary noises. With the development of deep learning technology, deep neural networks have shown excellent performance in processing complex and non-linear signals and have gradually been applied to the field of biological signal denoising.

[0004] Therefore, a real-time EEG signal denoising method based on deep neural network feedback control is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a real-time EEG signal denoising method based on deep neural network feedback control to solve the problems raised in the above background technique.

[0006] To achieve the above purpose, the present invention provides the following technical solution: A real-time EEG signal denoising method based on deep neural network feedback control, including the following steps:

[0007] Step 1: Collect the noisy EEG signal d[n] and the noise reference signal x[n] respectively through a composite electrode;

[0008] Step 2: Perform high-pass filtering on the noise reference signal x[n] to filter out low-frequency artifacts and retain high-frequency electromyographic noise components;

[0009] Step 3: Input the filtered noise reference signal x[n] into the deep neural network through a tapped delay line, and simultaneously input the noisy EEG signal d[n] into the deep neural network synchronously;

[0010] Step 4: The deep neural filter uses the deep neural network to model the noisy EEG signal d[n], generates a removal signal y[n] opposite to the noise signal, subtracts the noisy EEG signal d[n] from the removal signal y[n], and obtains the denoised EEG signal e[n];

[0011] Step Five: Apply the denoised EEG signal e[n] to the learning of the neural network, and the learning is completed through error backpropagation.

[0012] Preferably, the composite electrode in Step One includes a main EEG electrode and an auxiliary electrode. The main EEG electrode collects a noise reference signal x[n] at the Cz position, and the auxiliary electrode collects the noisy EEG signal d[n], and the obtained noise signal is separated for subsequent processing.

[0013] Preferably, the noise reference signal x[n] is processed by a high-pass filter to filter out low-frequency artifacts. The high-pass filter is a second-order Butterworth filter, and the frequency is set above 20 Hz to capture the spectral characteristics of muscle noise and other noises. The specific formula is as follows:

[0014]

[0015] where represents the second-order Butterworth high-pass filter for the outer electrode, γ represents the gain coefficient, is the second-order Butterworth high-pass filter for the inner electrode, BS[n] is the second-order Butterworth notch filter, LP ADC represents the low-pass filter for the sigma-delta converter.

[0016] Preferably, the deep neural network in Step Three is used to process the noise signal. The deep neural network uses a feedforward neural network with fully connected layers. The number of network layers is L = 6, and the number of neurons in each layer decreases to form a funnel structure. The output layer contains only one neuron, and the number of neurons is calculated as follows:

[0017]

[0018] where represents the number of taps in the delay line, is the base number of the previous layer calculated according to exponential decay.

[0019] Preferably, the activation function of the deep neural network in Step Three is the tanh function. The tanh function is linear at the origin and becomes non-linear as the signal strength increases. The tanh function learning can self-adjust non-linear processing, and the weights of the neurons are initialized to random values within the range of (0, 1). The formula for the forward propagation of the noise reference signal x[n] collected by the main EEG electrode through the first layer of the network is as follows:

[0020]

[0021] where Represents the activation value of the input layer of the neural network, is the weighted input value of the neuron, is the filtered signal of the j-th neuron and the k-th tap in the input layer, and x[n - k] represents the filtered signal of the k-th tap of the noise reference signal delay line.

[0022] Preferably, in the output layer of the entire neural network, the weighted sum generates the removal signal y[n], and the noise from the inner electrode can be cancelled, so as to obtain the denoised electroencephalogram signal e[n]. The specific formula for the weighted sum to generate the removal signal y[n] is as follows:

[0023]

[0024] Among them, represents the weighted input of the neuron in the output layer, is the weight between neurons in each layer, is the activation value of the i-th neuron in the penultimate layer;

[0025] The specific formula for the denoised electroencephalogram signal e[n] is as follows:

[0026] e[n]=d[n]-y[n].

[0027] Preferably, the denoised electroencephalogram signal e(n) finally output by the deep neural filter is also applied to the learning of the neural network through error backpropagation. The backpropagation error is calculated as follows:

[0028]

[0029] Among them, represents the error term of the j-th neuron in the layer, is the weight from the k-th neuron in the layer to the j-th neuron in the layer, is the derivative of the hyperbolic tangent function at ;

[0030] Preferably, the weight change of the best denoising effect of the deep neural filter is determined by the gradient descent rule, and the error signal updates the weight according to the gradient descent rule during the backpropagation process:

[0031]

[0032] Among them, η represents the learning rate, represents the activation value of the i-th neuron in the layer, represents the error term of the j-th neuron in the layer;

[0033] Among them, the effective learning rate in the deep neural network is proportional to the amplitude of the noise reference signal x[n]:

[0034]

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows: Through the innovative composite electrode design and the deep neural filter algorithm, the purity of the electroencephalogram (EEG) signal is significantly improved. The system uses the internal and external composite electrodes to collect the EEG signal and the noise reference signal, and realizes real-time adaptive noise cancellation through the deep learning algorithm, effectively removing electromyographic noise and artifacts, and improving the signal-to-noise ratio. The present invention not only has an efficient noise suppression effect, but also has strong environmental adaptability and computational efficiency, has stronger real-time adaptability and better denoising quality, is particularly suitable for real-time dynamic EEG signal denoising, and is suitable for intraoperative EEG signal monitoring for a long time and the application of portable EEG devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a schematic diagram of the principle of the intraoperative EEG signal real-time denoising method based on the deep neural network of the present invention;

[0037] Figure 2 is a structural diagram of the novel composite electrode of the present invention;

[0038] Figure 3 is a structural block diagram of the deep neural filter of the present invention;

[0039] Figure 4 is a comparison diagram of the signal power spectral density after being processed by the deep neural filter of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0041] Please refer to Figures 1 to 4 , the present invention provides a technical solution for a real-time EEG signal denoising method based on deep neural network feedback control:

[0042] A real-time EEG signal denoising method based on deep neural network feedback control includes the following steps:

[0043] Step 1: Collect the noisy EEG signal d[n] and the noise reference signal x[n] respectively through the composite electrode;

[0044] Step 2: Perform high-pass filtering on the noise reference signal x[n] to filter out low-frequency artifacts and retain high-frequency EMG noise components;

[0045] Step 3: Input the filtered noise reference signal x[n] into the deep neural network through a tapped delay line, and simultaneously input the noisy EEG signal d[n] into the deep neural network synchronously;

[0046] Step 4: The deep neural filter uses the deep neural network to model the noisy EEG signal d[n], generates a removal signal y[n] that is opposite to the noise signal, and subtracts the noisy EEG signal d[n] from the removal signal y[n] to obtain the denoised EEG signal e[n];

[0047] Step 5: Apply the denoised EEG signal e[n] to the learning of the neural network, and the learning is completed through error backpropagation;

[0048] The composite electrode in Step 1 includes a main EEG electrode and an auxiliary electrode. The main EEG electrode collects the noise reference signal x[n] at the Cz position, and the auxiliary electrode collects the noisy EEG signal d[n], and separates the obtained noise signal for subsequent processing;

[0049] The main EEG electrode and the auxiliary electrode are composed of two protruding annular parts, separated by a channel in the middle. The main EEG electrode is installed at the Cz position of the subject's head. The signal collected by the auxiliary electrode includes the EEG signal c[n] of interest, the background EEG activity b[n], and the noise m[n] caused by muscle activity, etc. Among them, the b[n] and m[n] signals form the total baseline noise r[n]. The main EEG electrode mainly collects the noise reference signal x[n] generated by artifacts, providing a data basis for subsequent noise cancellation. The specific formula is as follows:

[0050] r[n] = b[n] + m[n]

[0051] d[n] = r[n] + c[n]

[0052] x[n] = h[n] * (r[n] + α·c[n])

[0053] Where, r[n] represents the noise signal, c[n] represents the EEG signal of interest, h[n] represents the filter, and α represents the crosstalk coefficient.

[0054] First, the noise reference signal x[n] is high-pass filtered with a cut-off frequency above 20 Hz. The high-pass filter is a second-order Butterworth filter, which can maintain a flat frequency response and ensure that the removal of the low-frequency part does not affect the high-frequency noise characteristics, thereby enhancing the EMG noise. At the same time, the noisy EEG signal d[n] of the inner electrode is processed by DC filtering to remove unnecessary DC offsets, ensuring the stability of the input signal and providing a good baseline for subsequent processing;

[0055] The processing formulas are as follows:

[0056]

[0057] where γ represents the gain coefficient, represents the second-order Butterworth high-pass filter of the outer electrode, is the second-order Butterworth high-pass filter of the inner electrode, BS[n] is the second-order Butterworth notch filter, and LP ADC represents the low-pass filter for the sigma-delta converter;

[0058] The noise reference signal x[n] collected by the main EEG electrode is transmitted to the input layer of the deep neural network through a tapped delay line. The tapped delay line helps to provide temporal information for the network, especially to capture the impulse noise characteristics of EMG artifacts. The delay amount is:

[0059]

[0060] where f s represents the sampling frequency, is the cut-off frequency

[0061] At the same time, the noisy EEG signal d[n] collected by the auxiliary electrode is input through delay synchronization to ensure alignment with the noise reference signal, so that the network can learn the correlation features between the two. In step three, the deep neural network is used to process the noise signal. Through real-time learning, the deep neural network can effectively remove the common noise components in the inner and outer electrode signals while retaining the effective part of the EEG signal. The deep neural network uses a feedforward neural network with fully connected layers. The number of network layers is L = 6, and the number of neurons in each layer decreases, forming a funnel structure. The output layer contains only one neuron, and the number of neurons is calculated as follows:

[0062]

[0063] where, represents the number of taps in the delay line, is the base number of the previous layer calculated according to exponential decay;

[0064] In step 3, the activation function of the deep neural network is the tanh function. The tanh function is linear at the origin and becomes non-linear as the signal strength increases. The tanh function learning can self-adjust non-linear processing, and the function saturates when approaching ±1, thus preventing the vanishing gradient. The weights of the neurons are initialized to random values within the range of (0,1). The noise reference signal x[n] collected by the main EEG electrode is propagated forward through the first layer of the network by the following formula:

[0065]

[0066] where, represents the activation value of the input layer of the neural network, is the weighted input value of the neuron, is the filtered signal of the j-th neuron and the k-th tap in the input layer, and x[n-k] represents the filtered signal of the k-th tap of the noise reference signal delay line;

[0067] The network architecture adopts a multi-layer perceptron (MLP), and the activation function (tanh) is used between each layer. The output of each layer is obtained by the weighted sum of the previous layer, and its propagation formula in the network is as follows:

[0068]

[0069] represents the activation value of the j-th neuron in the l-th layer at time n, is the weighted input value of the j-th neuron in the l-th layer, is the weight between the i-th neuron in the l-th layer and the j-th neuron in the previous layer, is the activation value of the neurons in the (l-1)-th layer;

[0070] At the output layer of the entire neural network, the weighted sum generates the removal signal y[n], and the noise from the inner electrode can be cancelled, so as to obtain the denoised EEG signal e[n]. The specific formula for the weighted sum to generate the removal signal y[n] is as follows:

[0071]

[0072] where, represents the weighted input of the neurons in the output layer, is the weight between the neurons in each layer, is the activation value of the i-th neuron in the penultimate layer;

[0073] The specific formula for the denoised EEG signal e[n] is as follows:

[0074] e[n] = d[n] - y[n];

[0075] The denoised EEG signal e(n) finally output by the deep neural filter is also applied to the learning of the neural network through error backpropagation. The backpropagation error is calculated as follows:

[0076] δ L = e[n]

[0077]

[0078] where δ L represents the error signal of the neuron in the output layer (the L-th layer), represents the error term of the j-th neuron in the -th layer, is the weight from the k-th neuron in the -th layer to the j-th neuron in the -th layer, is the derivative of the hyperbolic tangent function at ;

[0079] The weight change for the best denoising effect of the deep neural filter is determined by the gradient descent rule. The error signal updates the weight according to the gradient descent rule during the backpropagation process:

[0080]

[0081] where η represents the learning rate, represents the activation value of the i-th neuron in the -th layer, represents the error term of the j-th neuron in the -th layer;

[0082] where the effective learning rate in the deep neural network is proportional to the amplitude of the noise reference signal x[n]:

[0083]

[0084] When the correlation between the noise reference signal x[n] and the error signal e[n] weakens, the learning process converges, meaning that the noise frequency components present in the outer ring electrodes will no longer appear in the output of the deep neural network, and thus the noise is removed.

[0085] To prevent overfitting and ensure robustness during the denoising process, the present invention designs a dynamic adjustment of the learning rate η eThe mechanism is that the adjustment of the learning rate is closely related to the amplitude change of the noise reference signal x[n]. The network automatically adjusts the learning rate according to the intensity and characteristics of the noise, so as to maintain a good denoising effect in different noise environments. Especially when the noise intensity changes greatly, the dynamic learning rate can effectively avoid the overfitting problem of the model and ensure that the network can adaptively adjust its parameters to keep the denoising effect in the best state at all times, as Figure 3 shown.

[0086] Finally, the Welch method is used to calculate the power spectral density of the inner electrode signal d[n] and the denoised EEG signal e[n], and the noise power in different frequency bands before and after denoising is obtained. The signal power is estimated by the P300 peak power. The signal-to-noise ratio is calculated according to the following formula, and the improvement of SNR before and after denoising is compared to verify the denoising effect of DNF;

[0087]

[0088] where v represents the following several types: the inner electrode signal, the output of the deep neural filter, the output of the FIR filter, the output of the Laplacian operator, and k is the index within the frequency range.

[0089] To verify the effectiveness of this method, the present invention adopts an experimental design based on P300 evoked potential to test the denoising effect of data from 10 subjects. In the experiment, the signal-to-noise ratio (SNR) of the inner electrode signal is significantly improved after being processed by the deep neural filter (DNF), indicating that this method has significant advantages in removing electromyogram artifact noise from EEG signals. Specifically, compared with the traditional LMS (Least Mean Square) adaptive filtering method, the deep neural network of the present invention can better retain the effective information of the EEG signal while removing electromyogram noise. Especially when removing large-amplitude electromyogram bursts, the network can automatically adjust its weights to achieve more accurate noise suppression, as Figure 4 shown.

[0090] By calculating the ratio of the signal power to the noise power (SNR) and comparing it with the SNR values before and after denoising, the denoising effect of the method of the present invention is further verified. The experimental results show that in the process of removing different types of noise, the method of the present invention can effectively improve the SNR of the EEG signal. Especially when the high-frequency electromyogram noise is strong, the deep neural network can dynamically adjust its processing strategy to ensure the stability of the denoising effect.

[0091] In addition, the present invention fully takes into account the individual differences among different subjects. In the experiment, randomly initialized network weights are adopted, and through personalized denoising processing methods, the EEG signals of each subject can obtain the optimal denoising effect. This personalized denoising strategy ensures that the deep neural filter can be adjusted according to the EEG characteristics and noise environment of different subjects, thereby improving the wide adaptability of the method in practical applications.

[0092] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time EEG signal denoising method based on deep neural network feedback control, characterized in that: The following steps are involved: Step 1: Collect the noisy EEG signal d[n] and the noisy reference signal x[n] respectively through the composite electrode; Step 2: Perform high-pass filtering on the noise reference signal x[n] to filter out low-frequency artifacts and retain high-frequency electromyographic noise components; Step 3: Input the filtered noise reference signal x[n] into the deep neural network through the tapped delay line, and simultaneously input the noisy EEG signal d[n] into the deep neural network; Step 4: The deep neural filter uses a deep neural network to model the noisy EEG signal d[n], generates a removal signal y[n] that is opposite to the noise signal, and subtracts the noisy EEG signal d[n] from the removal signal y[n] to obtain the denoised EEG signal e[n]; Step 5: Apply the denoised EEG signal e[n] to the learning of the neural network, and the learning is completed through error back propagation.

2. The method for real-time denoising of EEG signals based on deep neural network feedback control according to claim 1, characterized in that: The composite electrode in step 1 includes a main EEG electrode and an auxiliary electrode. The main EEG electrode collects a noise reference signal x[n] at the Cz position, and the auxiliary electrode collects a noisy EEG signal d[n], and separates the acquired noise signal for subsequent processing.

3. The method for real-time denoising of EEG signals based on deep neural network feedback control according to claim 1, characterized in that: The noise reference signal x[n] is processed by a high-pass filter to filter out low-frequency artifacts. The high-pass filter is a second-order Butterworth filter with a frequency set above 20 Hz to capture the spectral characteristics of muscle noise and other noises. The specific formula is as follows: Where γ represents the gain coefficient, represents an external electrode second-order Butterworth high-pass filter, is an inner electrode second-order Butterworth high-pass filter, BS[n] is a second-order Butterworth notch filter, LP ADC Represents a low-pass filter for a sigma-delta converter.

4. The method for real-time denoising of EEG signals based on deep neural network feedback control according to claim 1, characterized in that: In step 3, the deep neural network is used to process the noise signal. The deep neural network uses a feedforward neural network with a fully connected layer. The number of network layers is L=6. The number of neurons in each layer decreases to form a funnel structure. The output layer contains only one neuron. The number of neurons I(l) is calculated as follows: in, represents the number of taps in the delay line, b l-1 is the cardinality of the previous layer calculated according to exponential decay.

5. The method for real-time denoising of EEG signals based on deep neural network feedback control according to claim 1, characterized in that: The activation function of the deep neural network in step 3 is the tanh function. The tanh function is linear at the origin and becomes nonlinear as the signal strength increases. The tanh function learning can self-adjust the nonlinear processing. The weights of the neurons are initialized to random values ​​in the range of (0,1). The formula for the forward propagation of the noise reference signal x[n] collected by the main EEG electrode through the first layer of the network is as follows: in, represents the activation value of the neural network input layer, is the weighted input value of the neuron, is the filtered signal of the jth neuron and the kth tap in the input layer, and x[nk] represents the filtered signal of the kth tap of the noise reference signal delay line.

6. The method for real-time denoising of EEG signals based on deep neural network feedback control according to claim 4, characterized in that: In the output layer of the entire neural network, the weighted sum generates the removed signal y[n], and can offset the noise from the internal electrodes, thereby obtaining the denoised EEG signal e[n]. The specific formula for the weighted sum to generate the removed signal y[n] is as follows: in, represents the weighted input of the output layer neurons, is the weight between neurons in each layer, is the activation value of the i-th neuron in the second-to-last layer; The specific formula of the denoised EEG signal e[n] is as follows: e[n]=d[n]-y[n].

7. The method for real-time denoising of EEG signals based on deep neural network feedback control according to claim 6, characterized in that: The denoised EEG signal e(n) finally output by the deep neural filter is also applied to the learning of the neural network through error back propagation. The back propagation error is calculated as follows: in, represents the error term of the jth neuron in the lth layer, is the weight from the kth neuron in the l+1th layer to the jth neuron in the lth layer, is the hyperbolic tangent function in The derivative at .

8. The method for real-time denoising of EEG signals based on deep neural network feedback control according to claim 1, characterized in that: The weight change of the deep neural filter for the best denoising effect is determined by the gradient descent rule, and the error signal updates the weight according to the gradient descent rule during the back propagation process: Among them, η represents the learning rate, represents the activation value of the i-th neuron in the l-1th layer, represents the error term of the jth neuron in the lth layer; The effective learning rate in a deep neural network is proportional to the amplitude of the noisy reference signal x[n]:

Images