An input-aware steady-state regulation synaptic plasticity pulse convolutional neural network electroencephalogram emotion recognition method and system

Abstract

The present application belongs to the field of electroencephalogram emotion recognition and pulse neural network technology, and relates to a kind of input perception steady-state regulation's synaptic plasticity pulse convolutional neural network electroencephalogram emotion recognition method and system, method: after obtaining the multi-channel electroencephalogram signal of the subject to be identified, pre-processing is carried out and electroencephalogram feature sequence is constructed, electroencephalogram feature sequence is input to improved synaptic plasticity pulse convolutional neural network (synaptic plasticity pulse convolutional neural network is set in the input end of pulse convolution block Input perception steady-state regulation mechanism unit), obtain electroencephalogram emotion recognition result;System includes electroencephalogram data processing module, pulse convolution recognition module, input perception steady-state regulation mechanism unit, feedback modulation and synaptic plasticity updating module and output classification module.The method of the present application can enhance the heterogeneity of neurons, improve the performance of electroencephalogram emotion recognition;The system of the present application can improve the accuracy of electroencephalogram emotion recognition, and enhance the performance stability of the model on different subjects.

Description

Technical Field

[0001] This invention belongs to the fields of EEG emotion recognition, spiking neural networks and neuromorphic computing technology, and relates to an EEG emotion recognition method and system based on synaptic plasticity spiking convolutional neural networks with input perception steady-state regulation. Background Technology

[0002] Electroencephalography (EEG) is a type of physiological signal that records the electrical activity of the cerebral cortex in a non-invasive manner. It can directly reflect changes in neural activity during emotional induction, cognitive processing, and response to external stimuli. Compared to overt behavioral signals such as facial expressions, speech, and posture, EEG signals do not rely entirely on the individual's voluntary expression and have strong physiological objectivity and the ability to represent implicit emotions. Therefore, they are widely used in fields such as affective computing, human-computer interaction, brain-computer interfaces, mental health monitoring, and cognitive state assessment. However, how to extract stable and discriminative emotional features from EEG signals is a key technical problem in EEG emotion recognition.

[0003] Existing EEG emotion recognition methods typically consist of two stages: feature extraction and emotion classification. For feature extraction, mainstream methods often employ manually designed features such as power spectral density, differential entropy, and wavelet features to characterize emotion-related information in different frequency bands or channels. To further express the spatial distribution characteristics of EEG signals, some methods map multi-channel EEG features into two-dimensional feature maps according to the electrode space topology, or stack features from multiple frequency bands and time windows to construct a joint spatial, frequency, and temporal representation. These methods can enhance the expressive power of EEG features to some extent. For emotion classification, traditional EEG emotion recognition methods typically input manually extracted EEG features into classifiers such as support vector machines, random forests, K-nearest neighbors, or shallow neural networks for emotion category discrimination. These methods are relatively simple to implement, but their classification performance largely depends on the quality of feature design and manual parameter selection. In recent years, deep learning methods such as convolutional neural networks, recurrent neural networks, and graph neural networks have been widely used in EEG emotion recognition tasks. They can automatically learn high-level discriminative features in EEG signals through end-to-end training and improve recognition performance to some extent. However, traditional artificial neural networks typically employ continuous value activation and error backpropagation training mechanisms, which usually result in high computational energy consumption during model training and inference.

[0004] Spiking Neural Networks (SNNs), as a more bio-inspired neural network model, use discrete pulse sequences as information transmission carriers. They characterize the dynamic response of neurons through processes such as membrane potential accumulation, leakage, and threshold firing, exhibiting characteristics such as event-driven processing, sparse computation, low power consumption, and strong temporal information processing capabilities. Reference 1 (EESCN: A novel spiking neural network method for EEG-based emotion recognition[J]. Computer Methods and Programs in Biomedicine, 2024, 243:107927.) discloses a method for applying spiking neural networks to EEG emotion recognition, validating the application potential of SNNs in EEG emotion recognition tasks.

[0005] In spiking neural network training, spike-timing-dependent plasticity (STDP) is a typical local synaptic plasticity learning rule. STDP adjusts synaptic weights according to the relative timing of spike firing between presynaptic and postsynaptic neurons. When the presynaptic neuron fires before the postsynaptic neuron, the synaptic connection is strengthened, and vice versa. Reference 2 (Synaptic modifications in cultured hippocampal neurons: dependence on spike timing[J]. Journal of Neuroscience, 1998, 18(24): 10464-10472.) studied the physiological mechanism of STDP and confirmed that the timing of spikes has an important influence on synaptic regulation. In neural networks, trace-based STDP further represents the recent spike firing history of neurons by maintaining the decaying activity traces of presynaptic and postsynaptic neurons, and calculates the synaptic weight update based on the relationship between the current spike and the contralateral activity trace. Since the activity trace can be updated recursively with time steps, this method does not require matching all pre- and post-synaptic pulse pairs one by one, thus making it easy to implement in discrete-time spiking neural networks.

[0006] Existing STDP learning rules and their derivatives are essentially local learning rules, mainly adjusting weights based on the temporal relationship of pulses between adjacent neurons. However, in complex EEG emotion recognition tasks, relying solely on local STDP modulation processes lacks global task guidance such as error signals, which may lead to inconsistencies between the intermediate layer feature learning direction and the final classification target, thus affecting the model's discriminative ability. To address the problem of global error information transmission in deep network training, reference 3 (Random synaptic feedback weights support error backpropagation for deep learning[J]. Nature Communications, 2016, 7: 13276.) proposes a feedback alignment mechanism, which directly feeds back the output error signal to the hidden layer through fixed random feedback connections, thereby reducing the dependence on layer-by-layer accurate backpropagation and weight symmetry. This type of method provides a reference for introducing global task error information into local pulse plasticity learning.

[0007] Furthermore, reference 4 (Synaptic plasticity: taming the beast[J]. Nature Neuroscience, 2000, 3(Suppl): 1178–1183.) indicates that STDP-like Hebbian learning rules are prone to positive feedback effects during training. Specifically, neurons with strong responses become more active due to continuous synaptic enhancement, while neurons with weaker responses may gradually become silent, ultimately leading to an imbalance in network firing rate, feature expression degradation, and training instability. The homeostatic plasticity mechanism in neuroscience aims to maintain neuronal activity within a reasonable range through negative feedback regulation. Reference 5 (Cortical Motion Perception Emerges from Dimensionality Reduction with Evolved Spike-Timing-Dependent Plasticity Rules[J]. Journal of Neuroscience,2022, 42(30): 5882-5898.) discloses a spiking neural network learning rule that combines pulse timing-dependent plasticity and steady-state synaptic scaling. This method introduces a steady-state scaling term based on the average firing rate of postsynaptic neurons during STDP synaptic updates. Based on the deviation between the average firing rate of postsynaptic neurons and a preset target firing rate, the relevant synaptic weights are scaled and adjusted to maintain neuronal activity near the target firing level, thereby suppressing potential weight runaway and firing rate abnormalities during STDP training. However, this type of steady-state synaptic scaling method typically uses a preset fixed target firing rate as the steady-state setpoint, causing the long-term firing activity of different neurons to tend to converge at the preset target firing rate level. However, a fixed target firing rate may weaken the heterogeneity of neuronal activity and its adaptability to differences in input distribution.

[0008] Therefore, it is of great significance to study a method and system for EEG emotion recognition based on synaptic plasticity pulsed convolutional neural networks with input perception steady-state regulation in order to solve the problems existing in the prior art. Summary of the Invention

[0009] The technical problem to be solved by the present invention is that existing STDP-based spiking neural networks in EEG emotion recognition tasks suffer from problems such as lack of global task guidance in local learning, easy imbalance of neuron firing rate, overfiring of strong response neurons, silencing of weak response neurons, and the weakening of response heterogeneity by the fixed target firing rate steady-state mechanism. The present invention provides a method and system for EEG emotion recognition using a synaptic plasticity spiking convolutional neural network with input perception steady-state regulation.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] An input-sensory steady-state modulation synaptic plasticity spiking convolutional neural network brainwave emotion recognition method involves acquiring multi-channel brainwave signals from the subject to be identified, preprocessing them, constructing brainwave feature sequences, and inputting the brainwave feature sequences into an improved synaptic plasticity spiking convolutional neural network to obtain brainwave emotion recognition results.

[0012] The improved synaptic plasticity spiking convolutional neural network is a synaptic plasticity spiking convolutional neural network with an input-aware steady-state regulation mechanism unit set at the input end of the spiking convolution block. The spiking convolution block includes a convolutional layer, a LIF (Leaky Integrate-and-Fire) neuron layer, and an STDP synaptic plasticity local update mechanism unit based on DFA (Direct Feedback Alignment) modulation.

[0013] The input sensing steady-state regulation mechanism unit is connected to the convolutional layer, the LIF neuron layer, and the STDP synaptic plasticity local update mechanism unit based on DFA modulation.

[0014] The input perception steady-state regulation mechanism unit is used to acquire external stimulus information generated by the convolutional layer in the current pulse convolutional block based on the EEG feature sequence or output pulse feature of the previous layer (the previous layer refers to the network layer before the current pulse convolutional block; for the first pulse convolutional block, the previous layer is the EEG feature sequence; for the second and third pulse convolutional blocks, the previous layer is the pulse feature output of the previous level pulse convolutional block, respectively). It also acquires the pulse firing activity output by the LIF neuron layer in the current pulse convolutional block, statistically analyzes the long-term external stimulus and long-term firing state of the current layer neurons, generates an adaptive target firing rate based on the long-term external stimulus statistics, updates the steady-state gain based on the steady-state error between the adaptive target firing rate and the long-term firing state, and uses the steady-state gain to adjust the effective stimulus intensity input to the LIF neuron layer to generate the steady-state regulated pulse firing activity.

[0015] Existing steady-state modulation methods for spiking convolutional neural networks typically constrain neuron firing rate activity to a preset fixed value. While these methods can mitigate excessively strong or weak neuron activity to some extent, the fixed target value does not dynamically change with the feature strength of the input samples, data differences, or network layer responses. This can easily lead to over-homogenization of neuron responses under different input conditions. For high-information or strongly responding samples, the fixed firing rate constraint may suppress effective spiking activity, resulting in decreased feature representation ability; for low-information or weakly responding samples, it may introduce unnecessary firing enhancement, increasing noise response and the risk of misidentification. Therefore, this type of fixed-target modulation method struggles to balance sufficiency of expression with network stability under different input samples, limiting the model's classification performance and generalization ability in complex EEG emotion recognition tasks.

[0016] Therefore, how to avoid the problems of effective information suppression and noise amplification caused by uniformly constraining neuronal activity to a fixed firing rate or fixed target value in existing methods, and how to enable spiking neural networks to adaptively adjust steady-state control targets according to the characteristic states of different EEG input samples, has become a technical problem that urgently needs to be solved to improve the performance of EEG emotion recognition.

[0017] To address the aforementioned problems, this invention proposes an input-perception steady-state modulation synaptic plasticity spiking convolutional neural network method for EEG emotion recognition. Instead of simply unifying neuronal activity to a fixed target value, it constructs a steady-state gain based on the response state of the input EEG features, adaptively generating a steady-state modulation target that matches the current input, and dynamically constrains network spiking activity using synaptic plasticity mechanisms. This preserves effective spiking expression in strong response samples and suppresses invalid noise firing in weak response samples, achieving synergistic optimization between neuronal activity stability and feature expression ability. Consequently, it improves the accuracy, F1 score, and model generalization stability of positive, neutral, and negative EEG emotion recognition.

[0018] In STDP training, neurons with stronger responses are more likely to receive sustained synaptic enhancement, subsequently becoming more active and forming a positive feedback loop, leading to overfiring. Neurons with weaker responses, due to less firing and fewer opportunities for synaptic enhancement, may become increasingly less involved in feature expression, eventually approaching silencing. This invention dynamically adjusts the effective stimulus intensity input to the LIF neuronal layer through steady-state gain regulation. This aims to suppress overfiring and neuronal silencing while preserving the heterogeneity of responses of different neurons to different EEG inputs, thereby improving training stability and emotion classification and discrimination capabilities.

[0019] As a preferred technical solution:

[0020] As described above, the input-perception homeostasis-modulated synaptic plasticity spiking convolutional neural network EEG emotion recognition method involves the following processing steps performed by the input-perception homeostasis modulation mechanism unit:

[0021] (1) Obtain the intensity of the external stimulus output by the convolutional layer at the current time step. And based on the steady-state gain corresponding to the neuron Intensity of external stimuli Adjustments were made to obtain the effective stimulation intensity input to the LIF neuron layer. : ;

[0022] (2) The LIF neuron layer determines the effective stimulus intensity Update membrane potential and generate pulse firing activity: ;when hour, ;when hour, ;

[0023] in, This represents the membrane potential of the i-th neuron at time t. Indicates the membrane potential decay coefficient. Indicates the issuance threshold. This represents the pulse firing state of the i-th neuron at time t;

[0024] (3) Based on the pulse firing activity output by the LIF neuron layer, the long-term firing state of the corresponding neuron is updated using an exponential moving average method. ,satisfy: ;

[0025] in, The exponential moving average decay coefficient represents the long-term firing state and is used to control the influence of historical pulse firing states on the long-term firing state.

[0026] (4) Based on the intensity of the external stimulus output by the convolutional layer The long-term external stimulus statistics of the corresponding neurons are updated using an exponential moving average method. , Let represent the long-term external stimulus statistics of the i-th neuron at time t. satisfy: ;

[0027] in, The exponential moving average decay coefficient represents the long-term external stimulus statistics and is used to control the impact of historical external stimulus intensity on long-term external stimulus statistics.

[0028] (5) Based on long-term external stimulus statistics The external stimulus ratio is determined by the ratio between the external stimulus statistic and the mean long-term external stimulus statistic of neurons in the current layer involved in homeostasis regulation. ,satisfy: ;

[0029] in, This represents the average long-term external stimulus statistic of neurons in the current layer that participate in homeostasis regulation; N represents the number of neurons in the current layer that participate in homeostasis regulation.

[0030] (6) Based on the ratio of external stimuli Generate the adaptive target firing rate for the corresponding neuron ,satisfy: ;

[0031] in, Indicates the basic target distribution rate. Indicates the intensity of external stimulus modulation. and These represent the lower and upper bounds of the target distribution rate, respectively, and clip represents the cutoff function.

[0032] (7) Based on the adaptive target distribution rate With long-term distribution status steady-state error between Update steady-state gain Steady-state error satisfy: ;

[0033] To avoid frequent adjustments caused by small errors, the steady-state error is truncated based on a tolerance threshold to obtain the effective steady-state error; effective steady-state error Satisfy: When hour, ,when hour, ;

[0034] Steady-state gain satisfies: ;

[0035] in, Indicates the tolerance threshold. This indicates the steady-state gain adjustment step size. and These represent the lower and upper bounds of the steady-state gain, respectively.

[0036] (8) Utilize the updated steady-state gain Adjust the effective stimulation input to the LIF neuron layer at subsequent time steps.

[0037] As described above, in the input-sensory steady-state regulation synaptic plasticity pulse convolutional neural network EEG emotion recognition method, step (2) uses a soft reset method to update the membrane potential, that is, subtracting the threshold term while retaining some membrane potential information, rather than directly clearing the membrane potential to zero: .

[0038] The above-described input-perception homeostatic regulation synaptic plasticity pulse convolutional neural network EEG emotion recognition method acquires multi-channel EEG signals of the subject to be identified. Specifically, it collects multi-channel EEG signals of the subject during the emotion-evoking stimulation process. The emotion-evoking stimulation includes videos, pictures, sounds, and interactive tasks. The emotion categories include positive, neutral, and negative. In other application scenarios, emotion labels can also be represented by valence, arousal, or other emotion dimensions.

[0039] The above-described input-sensory steady-state modulation synaptic plasticity pulse convolutional neural network EEG emotion recognition method involves preprocessing and constructing an EEG feature sequence as follows: multi-channel EEG signals are sequentially subjected to bandpass filtering, power line interference suppression, removal of EEG or EMG artifacts, segmentation and normalization, and extraction of differential entropy features (DE features) within multiple time windows. The differential entropy features extracted within each time window are mapped to a two-dimensional spatial feature map according to the spatial topology of the EEG electrodes. The two-dimensional spatial feature maps corresponding to consecutive time windows are stacked according to the frequency and time dimensions to obtain a spatial-frequency-time EEG feature sequence. The size of the EEG feature sequence is represented as T×F×H×W, where T represents the number of consecutive time windows (i.e., time step), F represents the number of frequency bands, and H×W represents the size of the EEG electrode spatial topology grid.

[0040] The above-described input-perception steady-state modulation synaptic plasticity spiking convolutional neural network (SPL) EEG emotion recognition method comprises three cascaded spiking convolutional blocks and an output classification layer. Each of the three spiking convolutional blocks includes a convolutional layer, a LIF neuron layer, and a DFA-modulated STDP synaptic plasticity local update mechanism unit. Specifically, the convolutional layer of the first spiking convolutional block is an input encoding layer used to convert spatial-frequency-temporal EEG feature sequences into initial spiking feature maps. The second and third spiking convolutional blocks are hidden spiking convolutional blocks, and these hidden spiking convolutional blocks further include an input-perception steady-state modulation mechanism unit positioned between the convolutional feature extraction layer and the LIF neuron layer. The convolutional layer of the next spiking convolutional block uses the output spiking features of the previous layer as input to generate external stimuli for the current layer. The convolutional layers of the second and third spiking convolutional blocks serve as feature extraction layers. The output classification layer includes a global average pooling layer and a fully connected spiking classification layer, with the number of output nodes in the fully connected spiking classification layer matching the number of emotion categories.

[0041] The above-described input-sensory steady-state modulated synaptic plasticity spiking convolutional neural network (SPL) EEG emotion recognition method involves training the improved SPL before its application in EEG emotion recognition. The training process includes the following steps:

[0042] (1) Input the training set EEG feature sequence into the improved synaptic plasticity pulse convolutional neural network to obtain the training output emotion classification result;

[0043] (2) Calculate the classification error based on the emotion classification results and the corresponding emotion category labels;

[0044] (3) The output classification layer is trained for error correction based on the classification error, and a feedback modulation signal is generated;

[0045] (4) Use feedback modulation signal to modulate the STDP-based local update mechanism for synaptic plasticity to update the synaptic weights in each pulse convolution block;

[0046] (5) When the change in the value of the validation set loss function is lower than the preset convergence threshold during multiple rounds of training, the training is deemed complete; when the training completion condition is not met, the network parameters are iteratively updated based on the EEG feature sequence of the training set until the training completion condition is met or the preset maximum number of training rounds is reached.

[0047] The above-described input-sensory steady-state modulated synaptic plasticity pulsed convolutional neural network EEG emotion recognition method utilizes feedback modulation signals to modulate the local update mechanism of synaptic plasticity, specifically the synaptic weights connecting presynaptic neuron i and postsynaptic neuron j. Maintaining presynaptic activity traces and postsynaptic activity traces ,in , ;

[0048] in, and These represent the attenuation coefficients of the presynaptic and postsynaptic activity traces, respectively. and These represent the impulse states of the presynaptic neuron and the postsynaptic neuron at time t, respectively.

[0049] Based on the presynaptic and postsynaptic activity traces, the local weight changes without feedback modulation are first calculated and denoted as the STDP local candidate update quantities; the trace-based STDP local candidate update quantities satisfy: ;

[0050] in, and These represent the synaptic enhancement coefficient and the synaptic inhibition coefficient, respectively.

[0051] For the l-th pulse convolutional block, a fixed random feedback matrix is ​​used. Map the output error signal E to a global feedback modulation signal. , ;

[0052] The weight update of the l-th pulse convolutional block satisfies: ,in This indicates that the synaptic weights update the learning rate, and ⊙ represents element-wise multiplication. This represents the local candidate update matrix of the l-th pulse convolutional block obtained by trace-based STDP;

[0053] l=1,2,3.

[0054] Because standard STDP primarily learns rules locally, focusing only on the temporal relationship of pulses between neurons before and after the synapse, it can answer whether the connection between these two neurons should be strengthened or weakened recently, but it cannot directly answer whether this update is beneficial for the final emotion classification. Therefore, this invention introduces a global feedback modulation signal to add a "task direction" to the hidden layer STDP update. In this way, the weight update is simultaneously constrained by two types of information: one is the local pulse activity relationship, and the other is the global classification error signal. The ultimate goal is to improve the discriminative power of the hidden layer features and alleviate the problem of pure STDP lacking task guidance.

[0055] This invention also provides an input-aware steady-state modulated synaptic plasticity spiking convolutional neural network EEG emotion recognition system, used to perform the input-aware steady-state modulated synaptic plasticity spiking convolutional neural network EEG emotion recognition method as described above, comprising:

[0056] The EEG data processing module is used to acquire multi-channel EEG signals of the subject to be identified, preprocess the multi-channel EEG signals, and construct an EEG feature sequence.

[0057] The pulse convolution recognition module is used to input EEG feature sequences into an improved synaptic plasticity pulse convolutional neural network and output EEG emotion recognition results.

[0058] The input sensing steady-state regulation mechanism unit is used to obtain the external stimuli generated by the convolution feature extraction layer in the current pulse convolution block based on the output pulse features of the previous layer, obtain the pulse firing activity of the current LIF neuron layer, count the long-term external stimuli and long-term firing state of the current layer neurons, determine the adaptive target firing rate according to the external stimulus ratio, and update the steady-state gain according to the steady-state error.

[0059] The feedback modulation and synaptic plasticity update module is used to generate a feedback modulation signal based on the training output emotion classification result and emotion category label, and to use the feedback modulation signal to modulate the local candidate update amount of STDP to update the synaptic weights.

[0060] The output classification module is used to determine the emotion category corresponding to the EEG signal based on the output pulse, average firing rate, or output layer score.

[0061] The present invention also provides an electronic device or a computer-readable storage medium, the electronic device comprising a processor and a memory, the memory storing a computer program; or the computer-readable storage medium storing a computer program; wherein the computer program, when executed by the processor, implements the above-described input-perception steady-state regulation synaptic plasticity pulse convolutional neural network EEG emotion recognition method.

[0062] Beneficial effects:

[0063] (1) The present invention provides an input-sensory steady-state modulation of synaptic plasticity pulse convolutional neural network EEG emotion recognition method, which combines DFA global feedback modulation with trace-based STDP local update, so that the synaptic update of the pulse convolution block conforms to the pulse temporal relationship of pre- and post-synaptic neurons and has a global supervision direction related to the emotion classification target, thereby alleviating the problem of lack of task guidance in standard STDP.

[0064] (2) The input perception steady-state regulation of the present invention provides a method for EEG emotion recognition based on synaptic plasticity pulse convolutional neural network. The target firing rate is adaptively determined based on the long-term external stimulus statistics of neurons, so that neurons with stronger long-term external stimuli have a relatively higher target firing level and neurons with weaker long-term external stimuli have a relatively lower target firing level, thus avoiding all neurons from tending to the same fixed target firing rate.

[0065] (3) The input perception steady-state regulation of the present invention provides a method for EEG emotion recognition of synaptic plasticity pulse convolutional neural network. By dynamically adjusting the effective stimulation intensity entering the LIF neuron layer through steady-state gain, it can alleviate the problem of over-firing of strong response neurons and silencing of weak response neurons during STDP training, enhance the heterogeneity of neurons, and improve the EEG emotion recognition performance and training stability of pulse convolutional neural network.

[0066] (4) The input-sensory steady-state regulation synaptic plasticity pulse convolutional neural network EEG emotion recognition system of the present invention is applicable to EEG emotion datasets and can also be extended to other multi-channel EEG emotion recognition scenarios. Under the same EEG feature input, the same emotion category setting and the same training / validation division conditions, it can improve the category discrimination effect in the confusion matrix, improve the accuracy of EEG emotion recognition, and enhance the performance stability of the model on different subjects. Attached Figure Description

[0067] Figure 1 The original electroencephalogram (EEG) signal input for an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the overall model structure and training connection relationship of the pulse convolutional neural network with input-aware steady-state regulation provided in an embodiment of the present invention.

[0069] Figure 3 The confusion matrix comparison diagrams are shown for positive, neutral and negative EEG emotion recognition using a spiking convolutional neural network with and without an input-aware steady-state regulation mechanism on an EEG dataset. (a) is the baseline spiking convolutional neural network without the input-aware steady-state regulation mechanism, and (b) is the spiking convolutional neural network with the input-aware steady-state regulation mechanism of this invention.

[0070] Figure 4 The figure shows a comparison of ROC curves for EEG emotion recognition using different spiking convolutional neural networks on an EEG dataset. (a) is the ROC curve of the baseline spiking convolutional neural network without the input perception steady-state regulation mechanism, and (b) is the ROC curve of the spiking convolutional neural network after incorporating the input perception steady-state regulation mechanism of this invention. Detailed Implementation

[0071] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0072] A method for EEG emotion recognition based on synaptic plasticity and pulsed convolutional neural networks with input-aware steady-state regulation, comprising the following steps:

[0073] S1, Data Preprocessing;

[0074] During the emotional evoked stimulation process, multi-channel EEG signals of the subjects to be identified were collected. The multi-channel EEG signals were then subjected to bandpass filtering, power line interference suppression, removal of EEG or EMG artifacts, segmentation, and normalization. The emotional evoked stimuli included video, images, sound, and interactive tasks, and the emotion categories included positive, neutral, and negative.

[0075] S2. Constructing EEG feature sequences;

[0076] The EEG signal preprocessed in step S1 is segmented according to a preset time window. The differential entropy features extracted in each time window are mapped into a two-dimensional spatial feature map according to the spatial topology of the EEG electrodes. The two-dimensional spatial feature maps corresponding to the continuous time windows are stacked according to the frequency dimension and the time dimension to obtain the spatial-frequency-time EEG feature sequence. The size of the EEG feature sequence is represented as T×F×H×W, where T represents the number of continuous time windows, F represents the number of frequency bands, and H×W represents the size of the EEG electrode spatial topology grid.

[0077] S3. Construct an improved synaptic plasticity pulsed convolutional neural network;

[0078] The improved synaptic plasticity spiking convolutional neural network is a synaptic plasticity spiking convolutional neural network with an input-aware steady-state regulation mechanism unit set at the input end of the spiking convolutional block. The improved synaptic plasticity spiking convolutional neural network includes three cascaded spiking convolutional blocks and an output classification layer. Each of the three spiking convolutional blocks includes a convolutional feature extraction layer, a LIF neuron layer, and a STDP synaptic plasticity local update mechanism unit based on DFA modulation. Among them, the convolutional layer of the first spiking convolutional block is an input encoding layer, which is used to convert the spatial-frequency-temporal EEG feature sequence into an initial spiking feature map. The second and third spiking convolutional blocks are hidden spiking convolutional blocks, and the hidden spiking convolutional blocks further include an input-aware steady-state regulation mechanism unit set between the convolutional feature extraction layer and the LIF neuron layer. The convolutional feature extraction layer of the last spiking convolutional block uses the output spiking features of the previous layer as input and generates the external stimulus of the current layer.

[0079] The convolutional feature extraction layer is used to extract spatial, frequency, and temporal related features from the EEG feature sequence; the input perception homeostasis regulation mechanism unit is used to regulate the effective stimulation input to the LIF neuron layer in the corresponding hidden pulse convolution block; the LIF neuron layer converts the regulated continuous external stimulation (i.e., the effective stimulation finally obtained by the input perception homeostasis regulation mechanism unit) into a pulse sequence;

[0080] The input-sensing steady-state regulation mechanism unit is connected to the convolutional feature extraction layer, the LIF neuron layer, and the STDP synaptic plasticity local update mechanism unit based on DFA modulation.

[0081] The input sensing steady-state control mechanism unit performs the following processing steps:

[0082] (1) Obtain the intensity of external stimuli output by the convolutional feature extraction layer at the current time step. And based on the steady-state gain corresponding to the neuron Intensity of external stimuli Adjustments are made to obtain effective stimulation input to the LIF neuron layer. : ;

[0083] (2) The LIF neuron layer is based on effective stimulation Update membrane potential and generate pulse firing activity: ;when hour, ;when hour, The membrane potential is updated using a soft reset method, which involves subtracting a threshold term while retaining some membrane potential information, rather than directly resetting the membrane potential to zero. ;

[0084] in, This represents the membrane potential of the i-th neuron at time t. Indicates the membrane potential decay coefficient. Indicates the issuance threshold. This represents the pulse firing state of the i-th neuron at time t;

[0085] (3) Based on the pulse firing activity output by the LIF neuron layer, the long-term firing state of the corresponding neuron is updated using an exponential moving average method. ,satisfy: ;

[0086] in, The exponential moving average decay coefficient represents the long-term firing state and is used to control the influence of historical pulse firing states on the long-term firing state.

[0087] (4) The intensity of external stimuli output by the convolutional feature extraction layer The long-term external stimulus statistics of the corresponding neurons are updated using an exponential moving average method. , Let the long-term external stimulus statistic of the i-th neuron at time t satisfy: ;

[0088] in, The exponential moving average decay coefficient represents the long-term external stimulus statistics and is used to control the impact of historical external stimulus intensity on long-term external stimulus statistics.

[0089] (5) Based on long-term external stimulus statistics The external stimulus ratio of the i-th neuron is determined by the ratio between the external stimulus ratio and the average long-term external stimulus statistic of neurons in the current layer that participate in homeostasis. ,satisfy: ;

[0090] in, This represents the average long-term external stimulus statistic of neurons in the current layer involved in homeostasis regulation; N represents the number of neurons in the current layer involved in homeostasis regulation; when When, it indicates that the long-term external stimulus statistic of the i-th neuron is higher than the average level of the current layer; when When this occurs, it indicates that the long-term external stimulus statistics are lower than the current level.

[0091] (6) Based on the ratio of external stimuli Generate the adaptive target firing rate for the corresponding neuron ,satisfy: This formula allows neurons with strong long-term external stimulus statistics to have a relatively higher target firing rate, while neurons with weak long-term external stimulus statistics have a relatively lower target firing rate.

[0092] in, Indicates the basic target distribution rate. Indicates the intensity of external stimulus modulation. and These represent the lower and upper bounds of the target distribution rate, respectively, and clip represents the cutoff function.

[0093] (7) Based on the adaptive target distribution rate With long-term distribution status steady-state error between Update steady-state gain Steady-state error satisfy: When the long-term firing rate is lower than the adaptive target firing rate, the steady-state gain increases to enhance the effective stimulus; when the long-term firing rate is higher than the adaptive target firing rate, the steady-state gain decreases to suppress the effective stimulus.

[0094] The steady-state error is truncated based on a tolerance threshold to obtain the effective steady-state error; the effective steady-state error Satisfy: When hour, ,when hour, ;

[0095] Steady-state gain satisfies: ;

[0096] in, Indicates the tolerance threshold. This indicates the steady-state gain adjustment step size. and These represent the lower and upper bounds of the steady-state gain, respectively.

[0097] (8) Utilize the updated steady-state gain Adjust the effective stimulation input to the LIF neuron layer at subsequent time steps;

[0098] The output classification layer consists of a global average pooling layer and a fully connected spiking classification layer. The global average pooling layer is used to aggregate the spatial dimensions of the spiking feature map output by the last spiking convolutional block. The number of output nodes in the fully connected spiking classification layer is the same as the number of emotion categories, and error correction training is performed based on the classification error during the training phase. For example, when there are three emotion categories—positive, neutral, and negative—the number of output nodes is 3.

[0099] S4. The improved synaptic plasticity spiking convolutional neural network is trained before being used for EEG emotion recognition. The training process includes the following steps:

[0100] (1) Input the training set EEG feature sequence into the improved synaptic plasticity pulse convolutional neural network to obtain the training output emotion classification result;

[0101] (2) Calculate the classification error based on the emotion classification results and the corresponding emotion category labels;

[0102] (3) The output classification layer is trained for error correction based on the classification error, and a feedback modulation signal is generated, as follows:

[0103] An output error signal is generated based on the actual output of the output classification layer and the target label, and then processed by a fixed random feedback matrix. The output error signal is directly mapped to each pulse convolution block to form the global feedback modulation signal for the corresponding pulse convolution block. ;

[0104] Let O(t) be the actual output vector of the output classification layer at time t or within a sample time window, Y(t) be the target label vector, and E(t) be the output error vector. Then:

[0105] ;

[0106] For the l-th pulse convolutional block, a fixed random feedback matrix is ​​used. Map the output error signal E to a global feedback modulation signal. , Where l = 1, 2, 3;

[0107] This method does not rely on layer-by-layer chain-like backpropagation;

[0108] (4) Use feedback modulation signal to modulate the STDP-based local update mechanism for synaptic plasticity to update the synaptic weights in each pulse convolution block;

[0109] Synaptic weights connecting presynaptic neuron i and postsynaptic neuron j Maintain presynaptic activity traces respectively and postsynaptic activity traces ,in , ;

[0110] in, and These represent the attenuation coefficients of the presynaptic and postsynaptic activity traces, respectively. and These represent the impulse states of the presynaptic neuron and the postsynaptic neuron at time t, respectively.

[0111] Based on the presynaptic and postsynaptic activity traces, the local weight changes without feedback modulation are first calculated and denoted as the STDP local candidate update quantities; the trace-based STDP local candidate update quantities satisfy: ;

[0112] in, and These represent the synaptic enhancement coefficient and the synaptic inhibition coefficient, respectively. The above update avoids explicit enumeration of pulse pairs and is suitable for discrete-time pulse convolutional neural networks.

[0113] The weight update of the l-th pulse convolutional block satisfies: The formula shows that the DFA feedback signal mainly determines the task direction of weight update, while the STDP local candidate update quantity determines whether the weight update conforms to the local pulse timing relationship; where This indicates that the synaptic weights update the learning rate, and ⊙ represents element-wise multiplication. This represents the local candidate update matrix of the l-th pulse convolutional block obtained by trace-based STDP.

[0114] (5) When the change in the value of the validation set loss function is lower than the preset convergence threshold during multiple rounds of training, the training is deemed complete; when the training completion condition is not met, the network parameters are iteratively updated based on the EEG feature sequence of the training set until the training completion condition is met or the preset maximum number of training rounds is reached.

[0115] S5. Input the EEG feature sequence into the improved synaptic plasticity pulse convolutional neural network that has been trained. After input encoding by the first pulse convolutional block and pulse feature extraction by the second and third pulse convolutional blocks, the output classification layer obtains the output pulse, average firing rate or output layer score corresponding to each emotion category. The final EEG emotion recognition result is determined based on the firing rate vote, the maximum value of the average firing rate or the output layer score.

[0116] An input-perception steady-state modulated synaptic plasticity spiking convolutional neural network EEG emotion recognition system, used to execute the above-mentioned input-perception steady-state modulated synaptic plasticity spiking convolutional neural network EEG emotion recognition method, includes the following components:

[0117] The EEG data processing module is used to acquire multi-channel EEG signals of the subject to be identified, preprocess the multi-channel EEG signals, and construct an EEG feature sequence.

[0118] The pulse convolution recognition module is used to input EEG feature sequences into an improved synaptic plasticity pulse convolutional neural network and output EEG emotion recognition results.

[0119] The input sensing steady-state regulation mechanism unit is used to obtain the external stimuli generated by the convolution feature extraction layer in the current pulse convolution block based on the output pulse features of the previous layer, obtain the pulse firing activity of the current LIF neuron layer, count the long-term external stimuli and long-term firing state of the current layer neurons, determine the adaptive target firing rate according to the external stimulus ratio, and update the steady-state gain according to the steady-state error.

[0120] The feedback modulation and synaptic plasticity update module is used to generate a feedback modulation signal based on the training output emotion classification result and emotion category label, and to use the feedback modulation signal to modulate the local candidate update amount of STDP to update the synaptic weights.

[0121] The output classification module is used to determine the emotion category corresponding to the EEG signal based on the output pulse, average firing rate, or output layer score.

[0122] An electronic device or computer-readable storage medium, the electronic device including a processor and a memory, the memory storing a computer program; or the computer-readable storage medium storing a computer program; wherein the computer program, when executed by the processor, implements the above-mentioned input-perception steady-state regulation synaptic plasticity spiking convolutional neural network EEG emotion recognition method.

[0123] The following specific embodiments illustrate the electroencephalogram (EEG) emotion recognition method of the present invention, which utilizes synaptic plasticity pulsed convolutional neural network with input-aware steady-state regulation. Figure 2 As shown, the details are as follows:

[0124] This embodiment uses a publicly available EEG emotion dataset (SEED dataset) to verify the effectiveness of the method. This dataset may include multi-channel EEG signals collected from multiple subjects watching videos induced by positive, neutral, and negative emotions at multiple experimental stages; the specific number of subjects, number of channels, sampling rate, number of experimental stages, and duration of a single recording are subject to the description of the publicly available dataset used.

[0125] S1, Data Preprocessing;

[0126] For example Figure 1 The original EEG signal (multi-channel EEG signal) shown was bandpass filtered, downsampled to 200 Hz, and subjected to ICA (independent component analysis) artifact removal. Then, features of five frequency bands, namely delta wave, theta wave, alpha wave, beta wave, and gamma wave, were extracted.

[0127] S2. Constructing EEG feature sequences;

[0128] Based on the spatial topological relationship of EEG electrodes, the differential entropy features of each frequency band extracted within each time window are mapped into a two-dimensional spatial feature map. The features from multiple consecutive time windows are then concatenated to form an EEG feature sequence input to the pulse convolutional neural network. The size of the EEG feature sequence is represented as T×F×H×W, where T=6, F=5, and H×W=8×9.

[0129] S3. Construct an improved synaptic plasticity pulsed convolutional neural network;

[0130] The first pulse convolutional block uses a 3×3 kernel with 32 output channels, a stride of 1, and padding of 1. It outputs the initial pulse feature map through a LIF neuron layer. The second pulse convolutional block uses a 3×3 kernel with 64 output channels, a stride of 1, and padding of 1. The third pulse convolutional block uses a 3×3 kernel with 128 output channels, a stride of 1, and padding of 1. The convolutional outputs of the second and third pulse convolutional blocks are normalized by the corresponding input sensing steady-state regulation mechanism unit for effective stimulation modulation before being input into the LIF neuron layer to obtain the pulse feature map for the corresponding time step.

[0131] S4. An improved synaptic plasticity spiking convolutional neural network is trained before being used for EEG emotion recognition.

[0132] To ensure the stable convergence of the input sensing steady-state control mechanism, the membrane potential decay coefficient τ can be set to 0.95, and the firing threshold can be adjusted accordingly. Set to 1.0; Exponential moving average decay coefficient for long-term issuance status. and the exponential moving average decay coefficient of long-term external stimulus statistics All were set to 0.85; presynaptic activity trace attenuation coefficient and postsynaptic activity trace attenuation coefficient All are set to 0.80; the base target distribution rate is set to 0.15, and the lower bound of the target distribution rate is set to 0.80. Set to 0.05, upper bound of the target distribution rate. The external stimulus modulation intensity α is set to 0.7, the tolerance threshold ε is set to 0.0001, and the steady-state gain adjustment step size is set to 0.40. Set to 0.3, lower bound of steady-state gain Set to 0.1, upper bound of steady-state gain Set to 2.0. During training initialization, a long-term distribution state can be enabled. =0, Long-term external stimulus statistics =0, steady-state gain =1; the synaptic weight update learning rate η, synaptic enhancement coefficient η+, and synaptic inhibition coefficient η- can be set to 0.0001.

[0133] The above parameter ranges can be adjusted according to the size of the EEG dataset, the length of the time window, the number of network layers, and the number of emotion categories, and do not constitute a limitation on the scope of protection of this invention.

[0134] The input EEG feature sequence can include feature maps of multiple frequency bands, such as delta, theta, alpha, beta, and gamma bands, and can also include differential entropy, power spectral density, or time-frequency plot features. The output classification layer can be set to a three-class structure to output positive, neutral, and negative emotion categories, and can also be expanded to more categories depending on the application scenario.

[0135] The model results are primarily evaluated using the confusion matrix, ROC curve, and AUC metric. The confusion matrix reflects the misidentification rate among different emotion categories; the ROC curve represents the relationship between the true positive rate and the false positive rate at different classification thresholds; and the AUC, representing the area under the ROC curve, evaluates the model's classification ability. The true positive rate, false positive rate, and AUC metric satisfy the following formula:

[0136] ;

[0137] ;

[0138] ;

[0139] ;

[0140] ;

[0141] Where TP represents the number of true positive samples, FP represents the number of false positive samples, TN represents the number of true negative samples, FN represents the number of false negative samples, TPR represents the true positive rate, and FPR represents the false positive rate; Acc represents the accuracy, N_correct represents the number of correctly predicted samples, N_total represents the total number of samples included in the accuracy statistics, and F1 represents the harmonic mean of precision and recall, used to evaluate the overall performance of the model in category recognition.

[0142] The results of this experiment show that this invention can improve the recognition performance of various categories in EEG emotion recognition tasks, including positive, neutral, and negative emotions. Figure 3 As shown, Figure 3 In the diagram, (a) is the confusion matrix of the baseline pulsed convolutional neural network without the input-aware steady-state control mechanism, and (b) is the confusion matrix of the pulsed convolutional neural network after incorporating the input-aware steady-state control mechanism of this invention. Compared to... Figure 3In (a) and (b), the number of diagonal samples for positive, neutral, and negative categories all increased, while the number of misidentified off-diagonal samples decreased, indicating that the present invention can improve the classification accuracy between different emotion categories.

[0143] like Figure 4 As shown, the present invention also verifies the classification and discrimination ability using ROC curves. The average AUC of the baseline spiking convolutional neural network is approximately 0.868, while the average AUC of the synaptic plasticity spiking convolutional neural network with input perception steady-state regulation introduced in this invention is approximately 0.947, and the AUC of each emotion category is improved, indicating that the present invention can enhance the model's overall ability to distinguish between positive, neutral, and negative emotions.

[0144] Furthermore, under the same EEG feature input, the same training and verification division, and the same evaluation criteria, the model of this invention was compared with typical artificial neural network models and spiking neural network models. The experimental results are shown in Table 1.

[0145] Table 1. Comparison of experimental results of different models on the SEED dataset

[0146]

[0147] Where STD is the standard deviation, which is used to characterize the degree of dispersion of the recognition results of different subjects relative to the average recognition results;

[0148] As shown in Table 1, the IHASCNet (Input-Aware Steady-State Modulated Pulse Convolutional Network) corresponding to this invention achieves an accuracy of 83.10% and an F1 score of 78.80% on the SEED dataset, with corresponding standard deviations of 12.95 and 11.91, respectively. Overall, it outperforms IHASCNet-w / o IA without input-aware modulation and various comparative models, indicating that the input-aware steady-state modulation mechanism can improve classification performance while improving model stability.

Claims

1. A brainwave emotion recognition method based on synaptic plasticity pulsed convolutional neural networks with input-sensory steady-state regulation, characterized in that: After acquiring multi-channel EEG signals from the subject to be identified, preprocessing is performed and an EEG feature sequence is constructed. The EEG feature sequence is then input into an improved synaptic plasticity pulse convolutional neural network to obtain the EEG emotion recognition result. The improved synaptic plasticity pulsed convolutional neural network is a synaptic plasticity pulsed convolutional neural network with an input-sensing steady-state regulation mechanism unit set at the input end of the pulsed convolution block. The pulsed convolution block includes a convolutional layer, a LIF neuron layer, and a STDP synaptic plasticity local update mechanism unit based on DFA modulation. The input sensing steady-state regulation mechanism unit is connected to the convolutional layer, the LIF neuron layer, and the STDP synaptic plasticity local update mechanism unit based on DFA modulation. The input perception steady-state regulation mechanism unit is used to acquire external stimulus information generated by the convolutional layer in the current pulse convolutional block based on the EEG feature sequence or output pulse feature of the previous layer, and to acquire the pulse firing activity output by the LIF neuron layer in the current pulse convolutional block. It also collects statistics on the long-term external stimulus and long-term firing state of the neurons in the current layer, generates an adaptive target firing rate based on the long-term external stimulus statistics, updates the steady-state gain based on the steady-state error between the adaptive target firing rate and the long-term firing state, and uses the steady-state gain to adjust the effective stimulus intensity input to the LIF neuron layer to generate the steady-state regulated pulse firing activity. The input sensing steady-state control mechanism unit performs the following processing steps: (1) Obtain the intensity of the external stimulus output by the convolutional layer at the current time step. And based on the steady-state gain corresponding to the neuron Intensity of external stimuli Adjustments were made to obtain the effective stimulation intensity input to the LIF neuron layer. : ; (2) The LIF neuron layer determines the effective stimulus intensity Update membrane potential and generate pulse firing activity: ;when hour, ;when hour, ; in, This represents the membrane potential of the i-th neuron at time t. Indicates the membrane potential decay coefficient. Indicates the issuance threshold. This represents the pulse firing state of the i-th neuron at time t; (3) Based on the pulse firing activity output by the LIF neuron layer, the long-term firing state of the corresponding neuron is updated using an exponential moving average method. ,satisfy: ; in, The exponential moving average decay coefficient representing the long-term distribution status; (4) Based on the intensity of the external stimulus output by the convolutional layer The long-term external stimulus statistics of the corresponding neurons are updated using an exponential moving average method. ,satisfy: ; in, The exponential moving average decay coefficient representing the long-term external stimulus statistic. (5) Based on long-term external stimulus statistics The external stimulus ratio is determined by the ratio between the external stimulus statistic and the mean long-term external stimulus statistic of neurons in the current layer involved in homeostasis regulation. ,satisfy: ; in, This represents the average long-term external stimulus statistic of neurons in the current layer that participate in homeostasis regulation; N represents the number of neurons in the current layer that participate in homeostasis regulation. (6) Based on the ratio of external stimuli Generate the adaptive target firing rate for the corresponding neuron ,satisfy: ; in, Indicates the basic target distribution rate. Indicates the intensity of external stimulus modulation. and These represent the lower and upper bounds of the target distribution rate, respectively, and clip represents the cutoff function. (7) Based on the adaptive target distribution rate With long-term distribution status steady-state error between Update steady-state gain Steady-state error satisfy: ; The steady-state error is truncated based on a tolerance threshold to obtain the effective steady-state error; the effective steady-state error Satisfy: When hour, ,when hour, ; Steady-state gain satisfies: ; in, Indicates the tolerance threshold. This indicates the steady-state gain adjustment step size. and These represent the lower and upper bounds of the steady-state gain, respectively. (8) Utilize the updated steady-state gain The effective stimulation intensity input to the LIF neuron layer at subsequent time steps. Adjustments were made.

2. The EEG emotion recognition method based on synaptic plasticity pulsed convolutional neural networks with input-perception steady-state regulation according to claim 1, characterized in that, Step (2) Update the membrane potential using a soft reset method: .

3. The EEG emotion recognition method based on synaptic plasticity pulsed convolutional neural networks with input-perception steady-state regulation according to claim 2, characterized in that, The process involves acquiring multi-channel EEG signals from the subjects to be identified. Specifically, this involves collecting multi-channel EEG signals from the subjects during the process of evoking emotions. The evoking emotions include videos, pictures, sounds, and interactive tasks, and the emotion categories include positive, neutral, and negative.

4. The EEG emotion recognition method based on synaptic plasticity pulsed convolutional neural networks with input-perception steady-state regulation according to claim 3, characterized in that, Preprocessing and constructing the EEG feature sequence refers to: sequentially performing bandpass filtering, power frequency interference suppression, removal of electrooculogram (EOG) or electromyogram (EMG) artifacts, segmentation and normalization on multi-channel EEG signals, and extracting differential entropy features within multiple time windows; mapping the differential entropy features extracted within each time window to a two-dimensional spatial feature map according to the spatial topology of the EEG electrodes; stacking the two-dimensional spatial feature maps corresponding to consecutive time windows according to the frequency and time dimensions to obtain the spatial-frequency-time EEG feature sequence; the size of the EEG feature sequence is represented as T×F×H×W, where T represents the number of consecutive time windows, F represents the number of frequency bands, and H×W represents the size of the EEG electrode spatial topology grid.

5. The EEG emotion recognition method based on synaptic plasticity pulsed convolutional neural networks with input-perception steady-state regulation according to claim 4, characterized in that, The improved synaptic plasticity spiking convolutional neural network comprises three cascaded spiking convolutional blocks and an output classification layer. Each of the three spiking convolutional blocks includes a convolutional layer, a LIF neuron layer, and a STDP synaptic plasticity local update mechanism unit based on DFA modulation. Specifically, the convolutional layer of the first spiking convolutional block serves as an input encoding layer, used to convert spatial-frequency-temporal EEG feature sequences into initial spiking feature maps. The second and third spiking convolutional blocks are hidden spiking convolutional blocks, and the hidden spiking convolutional blocks further include an input perception steady-state regulation mechanism unit positioned between the convolutional layers and the LIF neuron layer. The convolutional layers of the second and third spiking convolutional blocks serve as feature extraction layers. The output classification layer includes a global average pooling layer and a fully connected spiking classification layer, with the number of output nodes in the fully connected spiking classification layer matching the number of emotion categories.

6. The EEG emotion recognition method based on synaptic plasticity pulsed convolutional neural networks with input-perception steady-state regulation according to claim 5, characterized in that, The improved synaptic plasticity spiking convolutional neural network was trained before being used for EEG emotion recognition. The training process included the following steps: (1) Input the training set EEG feature sequence into the improved synaptic plasticity pulse convolutional neural network to obtain the training output emotion classification result; (2) Calculate the classification error based on the emotion classification results and the corresponding emotion category labels; (3) The output classification layer is trained for error correction based on the classification error, and a feedback modulation signal is generated; (4) Use feedback modulation signal to modulate the STDP-based local update mechanism for synaptic plasticity to update the synaptic weights in each pulse convolution block; (5) When the change in the value of the validation set loss function is lower than the preset convergence threshold during multiple rounds of training, the training is deemed complete; when the training completion condition is not met, the network parameters are iteratively updated based on the EEG feature sequence of the training set until the training completion condition is met or the preset maximum number of training rounds is reached.

7. The EEG emotion recognition method based on synaptic plasticity pulsed convolutional neural networks with input-perception steady-state regulation according to claim 6, characterized in that, The STDP-based local update mechanism for synaptic plasticity is modulated using feedback modulation signals to adjust the synaptic weights connecting presynaptic neuron i and postsynaptic neuron j. Maintaining presynaptic activity traces and postsynaptic activity traces ,in , ; in, and These represent the attenuation coefficients of the presynaptic and postsynaptic activity traces, respectively. and These represent the impulse states of the presynaptic neuron and the postsynaptic neuron at time t, respectively. Based on the presynaptic and postsynaptic activity traces, the local weight changes without feedback modulation are first calculated and denoted as the trace-based STDP local candidate update quantity; the trace-based STDP local candidate update quantity satisfies: ; in, and These represent the synaptic enhancement coefficient and the synaptic inhibition coefficient, respectively. For the l-th pulse convolutional block, a fixed random feedback matrix is ​​used. Map the output error signal E to a global feedback modulation signal. , ; The weight update of the l-th pulse convolutional block satisfies: ,in This indicates that the synaptic weights update the learning rate, and ⊙ represents element-wise multiplication. This represents the local candidate update matrix of the l-th pulse convolutional block obtained by trace-based STDP; l=1,2,3。 8. An input-aware steady-state modulated synaptic plasticity spiking convolutional neural network EEG emotion recognition system, used to execute the input-aware steady-state modulated synaptic plasticity spiking convolutional neural network EEG emotion recognition method as described in any one of claims 1 to 7, characterized in that... include: The EEG data processing module is used to acquire multi-channel EEG signals of the subject to be identified, preprocess the multi-channel EEG signals, and construct an EEG feature sequence. The pulse convolution recognition module is used to input EEG feature sequences into an improved synaptic plasticity pulse convolutional neural network and output EEG emotion recognition results. The input sensing steady-state regulation mechanism unit is used to obtain the external stimuli generated by the convolutional layer in the current pulse convolution block based on the output pulse features of the previous layer, obtain the pulse firing activity of the current LIF neuron layer, count the long-term external stimuli and long-term firing state of the current layer neurons, determine the adaptive target firing rate according to the external stimulus ratio, and update the steady-state gain according to the steady-state error. The feedback modulation and synaptic plasticity update module is used to generate a feedback modulation signal based on the training output emotion classification result and emotion category label, and to use the feedback modulation signal to modulate the local candidate update amount of STDP to update the synaptic weights. The output classification module is used to determine the emotion category corresponding to the EEG signal based on the output pulse, average firing rate, or output layer score.

Images