Noninvasive electroencephalogram decoding characterization method of self-adaptive dynamic balance pulse neural network

Abstract

The invention belongs to the technical field of epileptic seizure prediction, and discloses a non-invasive electroencephalogram decoding characterization method of a self-adaptive dynamic balance pulse neural network, and the method specifically comprises the steps: a data preparation stage: carrying out the preprocessing purification of a non-invasive electroencephalogram signal, and obtaining a time window signal; a pulse neural network construction stage: coding the time window signal by using a classic pulse coding method to obtain a pulse sequence, and constructing a pulse neural network based on a multi-threshold pulse neuron model; in the electroencephalogram decoding characterization stage, a self-adaptive dynamic balance adjustment mechanism considering synaptic weight and network structure learning is used for carrying out optimization adjustment on local synaptic weight and global network parameters on the network, and non-invasive electroencephalogram decoding characterization is achieved. The invention aims to overcome the defect that the network is easy to fall into local optimum and cannot be quickly converged due to a learning mechanism, and fully improve the precision of network decoding characterization of the noninvasive electroencephalogram.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal decoding technology, specifically to a non-invasive EEG decoding and characterization method using an adaptive dynamic equilibrium spiking neural network. Background Technology

[0002] Electroencephalography (EEG) is a method of recording changes in electrical waves during brain activity using electrophysiological indicators. The development of brain decoding can not only promote the development of brain-computer interface (BCI) technology, but also effectively improve the daily lives of patients with neurological diseases from a clinical perspective.

[0003] In recent years, the application of deep learning to EEG decoding has made significant progress. Models based on convolutional neural networks and recurrent neural networks have demonstrated excellent performance in tasks such as motor imagery recognition, emotion recognition, and epileptic seizure prediction. However, most of these models are based on continuous-valued artificial neural networks, which have limited utilization of the rich spatiotemporal information in EEG signals and still have shortcomings in terms of energy consumption, computational cost, and biological interpretability. Spiking Neural Networks (SNNs), as the third generation of neural network models, use neuron models that mimic biological nervous systems for computational processing. Compared with the previous two generations of neural networks, SNNs have the ability to simulate a higher level of brain neural dynamics. However, the learning mechanism of network weights in existing research schemes is mostly based on the steady-state assumption of individual neurons, failing to fully consider the global regulation needs of neuronal clusters. It is difficult to accurately describe the complete dynamic process of neuronal cluster firing activity from steady-state disruption to reconstruction. Due to the insufficient characterization of neural dynamic regulation mechanisms, the network is prone to getting trapped in local optima and failing to converge quickly, thus limiting the accuracy of network decoding in representing non-invasive EEG.

[0004] Therefore, how to construct an adaptive dynamic balance adjustment mechanism that takes into account both synaptic weights and network structure learning, effectively alleviate the problems of slow convergence and local optima in the network, and thus improve the accuracy of non-invasive EEG signal decoding, is an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a non-invasive EEG decoding and representation method using an adaptive dynamic equilibrium spiking neural network to overcome the defect that the learning mechanism causes the network to easily get stuck in local optima and fail to converge quickly, so as to fully improve the accuracy of the network decoding and representation of non-invasive EEG.

[0006] To achieve the above objectives, this invention provides a non-invasive EEG decoding and characterization method using an adaptive dynamic equilibrium spiking neural network, comprising the following steps:

[0007] During the data preparation phase, non-invasive EEG signals are preprocessed and purified to obtain time window signals;

[0008] In the spiking neural network construction stage, the pulse sequence is obtained by encoding the time window signal using the classical pulse coding method, and the spiking neural network is constructed based on the multi-threshold spiking neuron model.

[0009] In the EEG decoding and representation stage, an adaptive dynamic balance adjustment mechanism that takes into account both synaptic weights and network structure learning is used to optimize and adjust the local synaptic weights and global network parameters of the network, thereby achieving non-invasive EEG decoding and representation.

[0010] In the spiking neural network construction phase, a multi-threshold spiking neuron model is used to construct the spiking neural network. The specific process of constructing the spiking neural network is as follows:

[0011] A multi-threshold LIF neuron model is constructed by using K LIF neuron models with different spiking activation thresholds:

[0012]

[0013]

[0014] In this system, each neuron defines multiple thresholds and maintains an equal interval distribution. Each threshold corresponds to a different pulse generation function. The thresholds of neurons at each level are configured using a fixed difference c, enabling neurons to trigger pulses under differentiated threshold conditions. The fixed difference c satisfies the following:

[0015]

[0016] in, k = 1, 2, ..., K represents the spiking threshold of the k-th neuron. This represents the pulse sequence of the k-th neuron in the n-th layer at time t. This represents the pulse sequence of all K neurons in the nth layer at time t.

[0017] A multi-threshold LIF neuron model is used to construct a spiking neural network, which is then used for nonlinear feature mapping of spiking sequences.

[0018]

[0019] Among them, O k (t) represents the output of the k-th layer neural network at time t, S k (t) represents the input received by the k-th layer of the neural network at time t. This represents the nonlinear mapping of the neuron model to the pulse sequence.

[0020] In the EEG decoding and representation stage, an adaptive dynamic balance adjustment mechanism that takes into account both synaptic weights and network structure learning is used to optimize and adjust the local synaptic weights and global network parameters of the network, thereby achieving non-invasive EEG decoding and representation.

[0021] The "adaptive" nature of the non-invasive EEG decoding representation method is based on a synaptic weight optimization mechanism. This mechanism modifies synaptic weights through local steady-state learning to dynamically adapt to synaptic input. Specifically, it optimizes the parameters of the spiking neuron's membrane capacitance and resistance through local steady-state learning, enabling the neuron to maintain an appropriate pulse firing rate. The synaptic weight optimization mechanism is described as follows:

[0022]

[0023] Where, τ ip This represents the relative integration rate between the neuron model and the internal synaptic plasticity mechanism. Let represent the reciprocals of membrane capacitance and membrane resistance, respectively; β represent the scaling factor; I(t) represent the presynaptic input current of the neuron; O(t) represent the output response of the neuron; ∈ represents the impulse intensity; δ(tt) ( f)) represents the Dirac function, expressed as at time t ( f) is the neuron impulse value.

[0024] The "dynamic equilibrium" of the non-invasive EEG decoding representation method is based on a network structure learning mechanism, consisting of a direct feedback alignment algorithm and a neuronal population dynamics adjustment mechanism. This optimizes the global network parameters to improve the noise resistance and representation accuracy of the decoding.

[0025] Specifically, a fixed random feedback matrix is ​​used. The error signal of the output layer neuron is passed through a random alignment vector b in B. jk The signal is directly propagated to hidden layer neuron j, yielding the local error signal of the hidden layer. Its update rule can be expressed as:

[0026]

[0027] in, This represents the error signal of the hidden layer neurons. The error signal of the output layer neuron is defined as the value of the random alignment vector from output layer neuron k to hidden layer neuron j. denoted as the synaptic input current of hidden layer neuron j at time t.

[0028] After obtaining the error signal of the hidden layer, the weight updates from the input layer to the hidden layer and from the hidden layer to the output layer are expressed as follows:

[0029]

[0030] Where, Δw ij Δw represents the weight update values ​​of input layer neuron i and hidden layer neuron j. jk η represents the weight update values ​​of hidden layer neuron j and output layer neuron k, and η represents the learning rate. This represents the pulse sequence function of neuron i in the input layer. This represents the pulse sequence function of hidden layer neuron j.

[0031] Furthermore, based on the neuron population dynamics mechanism, the average firing rate of the neuron population is updated to optimize the network structure:

[0032]

[0033] Where E,I represents the discharge rate, h E ,h I Indicates an external constant, ε E ,ε I Indicates the time scale parameter. This indicates the strength of connections between groups of excitatory neurons. This indicates the connection strength between excitatory and inhibitory neuronal populations. This indicates the connection strength between inhibitory neuron populations and excitatory neuron populations. ζ represents the connection strength between the inhibitory neuron population itself, ζ represents the excitation response function, and D represents the time delay.

[0034] During network training, the loss function L is defined as the mean square error between the network output and the corresponding label, which can be expressed as:

[0035]

[0036] stW * (α)=argmin(L t (A t ;W,α)+μL ip (W,α))

[0037] Where N represents the input sample size, R i O represents the label of the i-th sample. i Indicates network output, A t A represents the training phase data. v L represents data from the testing phase. t L represents the expected loss during the training phase. v L represents the loss during the testing phase. ipLet Ω(·) represent the local loss, λ represent the penalty term, λ represent the regularization strength coefficient, W represent the model weight parameters, α represent the model hyperparameters, and μ represent the local weight parameters during the training phase. * (α) represents minimizing the expected training loss L on the training set given the hyperparameter α. t With local loss L ip The obtained optimal model weights.

[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a non-invasive EEG decoding and characterization method using an adaptive dynamic equilibrium spiking neural network, as described in this invention.

[0040] Figure 2 This is a flowchart illustrating the steps of the spiking neural network construction method according to an embodiment of the present invention;

[0041] Figure 3 This is a flowchart illustrating the adaptive dynamic balance adjustment mechanism according to an embodiment of the present invention;

[0042] Figure 4 This is a network structure diagram of the adaptive dynamic equilibrium spiking neural network of the present invention. Detailed Implementation

[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] like Figure 1 As shown, this invention is a non-invasive EEG decoding and characterization method using an adaptive dynamic equilibrium spiking neural network, comprising the following steps:

[0045] During the data preparation phase, non-invasive EEG signals are preprocessed and purified to obtain time window signals;

[0046] In the spiking neural network construction stage, the pulse sequence is obtained by encoding the time window signal using the classical pulse coding method, and the spiking neural network is constructed based on the multi-threshold spiking neuron model.

[0047] In the EEG decoding and representation stage, an adaptive dynamic balance adjustment mechanism that takes into account both synaptic weights and network structure learning is used to optimize and adjust the local synaptic weights and global network parameters of the network, thereby achieving non-invasive EEG decoding and representation.

[0048] Example

[0049] This embodiment is based on a non-invasive EEG decoding and representation method using an adaptive dynamic equilibrium spiking neural network, which mainly includes the following steps: data preparation stage, spiking neural network construction stage, and EEG decoding and representation stage.

[0050] In this embodiment, the data preparation stage involves preprocessing and purifying the non-invasive EEG signal to obtain a time window signal. The preprocessing operations include filtering and time window segmentation. Specifically, the filtering operation uses a bandpass filter to limit the signal frequency to 0.5 to 64 Hz, followed by notch filtering to eliminate 50 Hz power frequency noise. The time window segmentation operation divides the non-invasive EEG signal into time windows of length T to obtain a shape (B, C, T × f). s The time window signal, where B is the number of samples, C is the number of electrode channels for non-invasive EEG signals, and f s This represents the sampling rate of non-invasive EEG signals. Optionally, a certain overlap rate can be set between signals within a time window, which can reflect the correlation between signals to some extent.

[0051] In this embodiment, during the construction stage of the spiking neural network, the time window signal is encoded using the classical pulse coding method to obtain a pulse sequence, and the spiking neural network is constructed based on the multi-threshold spiking neuron model.

[0052] Optionally, the encoding step of the time window signal using the classical pulse coding method can employ methods such as Time-To-First-Spike (TTFS) or Direct Input Coding (DIC). For example, in TTFS, the amplitude of the time window signal is mapped to the pulse occurrence time, with higher amplitude values ​​corresponding to earlier pulse moments, thus expressing amplitude information on the time axis.

[0053] like Figure 2 As shown, in the spiking neural network construction stage, a multi-threshold spiking neuron model is used to construct the spiking neural network, which includes an input layer, a hidden layer, and an output layer. The input layer is used to receive the pulse sequence after pulse encoding, the hidden layer is used to simulate biological brain neurons, and the output layer is used to perform classification decisions. The specific process of constructing the spiking neural network is as follows:

[0054] A multi-threshold LIF neuron model is constructed from K LIF neuron models with different spiking activation thresholds:

[0055]

[0056] In this system, each neuron defines multiple thresholds and maintains an equal interval distribution. Each threshold corresponds to a different pulse generation function. The thresholds of neurons at each level are configured using a fixed difference c, enabling neurons to trigger pulses under differentiated threshold conditions. The fixed difference c satisfies the following:

[0057]

[0058] in, k = 2, ..., K represents the spiking threshold of the k-th neuron. This represents the pulse sequence of the k-th neuron in the n-th layer at time t. This represents the pulse sequence of all K neurons in the (n-1)th layer at time t.

[0059] A multi-threshold LIF neuron model is used to construct a spiking neural network, which is then used for nonlinear feature mapping of spiking sequences.

[0060]

[0061] Among them, O k (t) represents the output of the k-th layer neural network at time t, S k (t) represents the input received by the k-th layer of the neural network at time t. This represents the nonlinear mapping of the pulse sequence to the multi-threshold LIF neuron model.

[0062] like Figure 3 As shown, in this embodiment, the EEG decoding and representation stage utilizes an adaptive dynamic balance adjustment mechanism that balances synaptic weights and network structure learning to optimize and adjust local synaptic weights and global network parameters, thereby achieving non-invasive EEG decoding and representation, as detailed below:

[0063] The "adaptive" nature of the non-invasive EEG decoding representation method is based on a synaptic weight optimization mechanism. This mechanism modifies synaptic weights through local steady-state learning to dynamically adapt to synaptic input. Specifically, it optimizes the parameters of the spiking neuron's membrane capacitance and resistance through local steady-state learning, enabling the neuron to maintain an appropriate pulse firing rate. The synaptic weight optimization mechanism is described as follows:

[0064]

[0065] Where, τ ip This represents the relative integration rate between the neuron model and the internal synaptic plasticity mechanism. Let represent the reciprocals of membrane capacitance and membrane resistance, respectively; β represent the scaling factor; I(t) represent the presynaptic input current of the neuron; O(t) represent the output response of the neuron; ∈ represents the impulse intensity; δ(tt) (f)) represents the Dirac function, expressed as at time t ( f) is the neuron impulse value.

[0066] The non-invasive EEG decoding representation method, "Dynamic Equilibrium," is based on a network structure learning mechanism. It consists of a direct feedback alignment algorithm and a neuronal population dynamics adjustment mechanism, which optimizes the global network parameters to improve the noise resistance and representation accuracy of the decoding.

[0067] Specifically, a randomized fixed feedback matrix is ​​used. The error signal of the output layer neuron is passed through a random alignment vector b. jk The signal is directly propagated to hidden layer neuron j, yielding the local error signal of the hidden layer. Its update rule can be expressed as:

[0068]

[0069] in, This represents the error signal of the hidden layer neurons. b represents the error signal of the output layer neuron. jk The element in the random feedback matrix B is defined as the value of the random alignment vector from output layer neuron k to hidden layer neuron j. denoted as the synaptic input current of hidden layer neuron j at time t.

[0070] After obtaining the error signal of the hidden layer, the weight updates from the input layer to the hidden layer and from the hidden layer to the output layer are expressed as follows:

[0071]

[0072] Where, Δw ij Δw represents the weight update values ​​of input layer neuron i and hidden layer neuron j. jk η represents the weight update values ​​of hidden layer neuron j and output layer neuron k, and η represents the learning rate. This represents the pulse sequence function of neuron i in the input layer. This represents the pulse sequence function of hidden layer neuron j.

[0073] Furthermore, based on the neuron population dynamics mechanism, the average firing rate of the neuron population is updated to optimize the network structure:

[0074]

[0075] Where E,I represents the discharge rate, h E ,h I Indicates an external constant, ε E ,ε I Indicates the time scale parameter. This indicates the connection strength (self-excitation) between groups of excitatory neurons. This indicates the connection strength between excitatory and inhibitory neuronal populations (excitation drives inhibition). This indicates the strength of the connection between inhibitory neuron populations and excitatory neuron populations (inhibitory feedback). ζ represents the connection strength (self-inhibition) between inhibitory neuron populations, ζ represents the excitation response function, and D represents the time delay.

[0076] During network training, the loss function L is defined as the mean square error between the network output and the corresponding label, which can be expressed as:

[0077]

[0078] stW * (α)=argmin(L t (A t ;W,α)+μL ip (W,α))

[0079] Where N represents the input sample size, R i O represents the label of the i-th sample. i Indicates network output, A t A represents the training phase data. v L represents data from the testing phase. t L represents the expected loss during the training phase. v L represents the loss during the testing phase. ip Let Ω(·) represent the local loss, λ represent the penalty term, λ represent the regularization strength coefficient, W represent the model weight parameters, α represent the model hyperparameters, and μ represent the local weight parameters during the training phase. * (α) represents minimizing the expected training loss L on the training set given the hyperparameter α. t With local loss L ip The obtained optimal model weights.

[0080] Therefore, this invention employs the aforementioned adaptive dynamic balance spiking neural network EEG decoding method. The proposed multi-threshold mechanism enables neurons to generate pulses at different activation levels, capturing subtle changes in input signals and thus enhancing the network's representational ability. The proposed adaptive dynamic balance adjustment mechanism, which balances synaptic weights and network structure learning, can modify synaptic weights from local steady-state learning, maintaining the neuronal internal firing activity at a steady-state level, thereby maximizing the information entropy of the neuron's output. Global learning can directly propagate the output layer error signal to the hidden layer, mitigating the weight transmission attenuation problem in error backpropagation, thereby significantly improving the accuracy of network decoding representation of non-invasive EEG.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate preferred examples of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A non-invasive EEG decoding and characterization method using an adaptive dynamic equilibrium spiking neural network, characterized in that: During the data preparation phase, non-invasive EEG signals are preprocessed and purified to obtain time window signals; In the spiking neural network construction stage, the pulse sequence is obtained by encoding the time window signal using the classical pulse coding method, and the spiking neural network is constructed based on the multi-threshold spiking neuron model. In the EEG decoding and representation stage, an adaptive dynamic balance adjustment mechanism that takes into account both synaptic weights and network structure learning is used to optimize and adjust the local synaptic weights and global network parameters of the network, thereby achieving non-invasive EEG decoding and representation.

2. The non-invasive EEG decoding and characterization method of an adaptive dynamic equilibrium spiking neural network according to claim 1, characterized in that, The spiking neural network is constructed from K LIF neuron models with different spiking activation thresholds, forming a multi-threshold LIF neuron model: In this system, each neuron defines multiple thresholds and maintains an equal interval distribution. Each threshold corresponds to a different pulse generation function. The thresholds of neurons at each level are configured using a fixed difference c, enabling neurons to trigger pulses under differentiated threshold conditions. The fixed difference c satisfies the following: in, This represents the firing threshold of the k-th neuron. This represents the pulse sequence of the k-th neuron in the n-th layer at time t. This represents the pulse sequence of all K neurons in the nth layer at time t. Furthermore, a spiking neural network is constructed based on a multi-threshold LIF neuron model to perform nonlinear feature mapping on the spiking sequence: Among them, O k (t) represents the output of the k-th layer neural network at time t, S k (t) represents the input received by the k-th layer of the neural network at time t. This represents the nonlinear mapping of the neuron model to the pulse sequence.

3. The method according to claim 1, characterized in that, The adaptive dynamic balance adjustment mechanism that takes into account both synaptic weights and network structure learning is an optimized adjustment of local synaptic weights and global network parameters to achieve non-invasive EEG decoding representation. The "adaptive" nature of the non-invasive EEG decoding representation method is based on a synaptic weight optimization mechanism. This mechanism modifies synaptic weights through local steady-state learning to dynamically adapt to synaptic input. Specifically, it optimizes the parameters of the spiking neuron's membrane capacitance and resistance through local steady-state learning, enabling the neuron to maintain an appropriate pulse firing rate. The synaptic weight optimization mechanism is described as follows: Where, τ ip This represents the relative integration rate between the neuron model and the internal synaptic plasticity mechanism. Let represent the reciprocals of membrane capacitance and membrane resistance, respectively; β represent the scaling factor; I(t) represent the presynaptic input current of the neuron; O(t) represent the output response of the neuron; ∈ represents the impulse intensity; t(tt) (f) ) represents the Dirac function, expressed as at time t (f) The neuronal impulse value. The "dynamic equilibrium" of the non-invasive EEG decoding representation method is based on a network structure learning mechanism, consisting of a direct feedback alignment algorithm and a neuronal population dynamics adjustment mechanism. This optimizes the network parameters globally to improve the noise resistance and representation accuracy of the decoding. Specifically, a fixed random feedback matrix is ​​used. The error signal of the output layer neuron is passed through a random alignment vector b in B. jk The signal is directly propagated to hidden layer neuron j, yielding the local error signal of the hidden layer. Its update rule can be expressed as: in, This represents the error signal of the hidden layer neurons. The error signal of the output layer neuron is defined as the value of the random alignment vector from output layer neuron k to hidden layer neuron j. denoted as the synaptic input current of hidden layer neuron j at time t. After obtaining the error signal of the hidden layer, the weight updates from the input layer to the hidden layer and from the hidden layer to the output layer are expressed as follows: Where, Δw ij Δw represents the weight update values ​​of input layer neuron i and hidden layer neuron j. jk η represents the weight update values ​​of hidden layer neuron j and output layer neuron k, and η represents the learning rate. This represents the pulse sequence function of neuron i in the input layer. This represents the pulse sequence function of hidden layer neuron j. Furthermore, based on the neuron population dynamics mechanism, the average firing rate of the neuron population is updated to optimize the network structure: Where E,I represents the discharge rate, h E ,h I Indicates an external constant, ε E ,ε I Indicates the time scale parameter. This indicates the strength of connections between groups of excitatory neurons. This indicates the connection strength between excitatory and inhibitory neuronal populations. This indicates the connection strength between inhibitory neuron populations and excitatory neuron populations. ζ represents the connection strength between the inhibitory neuron population itself, ζ represents the excitation response function, and D represents the time delay. During network training, the loss function L is defined as the mean square error between the network output and the corresponding label, which can be expressed as: stW * (a)=argmin(L t (A t ;W,a)+μL ip (W,a)) Where N represents the input sample size, R i O represents the label of the i-th sample. i Indicates network output, A t A represents the training phase data. v L represents data from the testing phase. t L represents the expected loss during the training phase. v L represents the loss during the testing phase. ip Let Ω(·) represent the local loss, λ represent the penalty term, λ represent the regularization strength coefficient, W represent the model weight parameters, α represent the model hyperparameters, and μ represent the local weight parameters during the training phase. * (α) represents minimizing the expected training loss L on the training set given the hyperparameter α. t With local loss L ip The obtained optimal model weights.

Images