Nerve stress injury identification algorithm based on multi-attention crossover
By constructing an end-to-end differentiable joint optimization framework, and combining adaptive signal preprocessing and multi-dimensional attention feature fusion with dynamic spiking neural networks, the adaptive and collaborative optimization problems in identifying complex neural stress injuries in traditional methods are solved, achieving high-precision injury identification and localization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-24
- Publication Date
- 2026-03-13
AI Technical Summary
Existing technologies struggle to efficiently and accurately identify complex neural stress injury patterns. Traditional methods suffer from poor adaptability, fixed preprocessing parameters, simple feature fusion methods, and fragmented model modules that cannot be collaboratively optimized.
An end-to-end differentiable joint optimization framework is constructed, which combines adaptive signal preprocessing, multi-dimensional attention feature fusion and dynamic spiking neural network decision mechanism. Through learnable parameter optimization filtering, EMD decomposition, multi-attention fusion and LIF neural network, adaptive enhancement and multimodal interaction of EEG signals are achieved.
It significantly improves the ability to capture subtle damage features, enhances the model's individual adaptability and transparency, provides high-precision damage identification and localization support, and assists in clinical diagnosis.
Smart Images
Figure CN121647599A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of artificial intelligence and biomedical engineering, and in particular relates to a neural stress injury recognition algorithm based on multi-attention crossover. Background Technology
[0002] Electroencephalography (EEG) is a non-invasive technique that assesses brain function by recording the electrical activity of neurons in the brain. It is of great value in identifying neurological stress (such as post-traumatic stress disorder) and brain injury. However, EEG signals are very weak and easily affected by interference from eye movements, muscle activity, etc. (i.e., "artifacts"), and signal patterns vary from person to person and change constantly, which makes accurate identification of injuries very difficult.
[0003] Traditional analysis methods rely on fixed filtering and manual feature extraction, which are cumbersome, poorly adaptable, and struggle to capture weak damage signals. In recent years, deep learning models (such as CNNs and RNNs) have been used to automatically extract EEG features, improving recognition performance. However, these models are energy-intensive, differ significantly from the actual brain's working mechanism, and their decision-making processes are opaque. Current research attempts to combine SNNs with attention mechanisms or improve SNN neuron structures, but significant shortcomings remain: preprocessing parameters are fixed and cannot be adaptively optimized; feature fusion methods are simplistic and fail to fully utilize multi-dimensional information such as time, space, and frequency; and the overall process is fragmented, with modules unable to evolve collaboratively.
[0004] Therefore, existing technologies struggle to efficiently and accurately identify complex neural injury patterns. This invention aims to address these issues by proposing a novel, learnable, and adaptively optimized algorithm that enables high-precision, interpretable identification of neural stress injuries throughout the entire process, from signal preprocessing to final decision-making. Summary of the Invention
[0005] The purpose of this invention is to provide a neural stress injury recognition algorithm based on multi-attention crossover. By constructing an end-to-end differentiable joint optimization framework, it deeply integrates adaptive signal preprocessing, multi-dimensional attention feature fusion, and dynamic spiking neural network decision-making mechanism. This achieves adaptive enhancement of weak injury features in EEG signals, multimodal interaction, and brain-like recognition. It solves the problems in the prior art, such as fixed preprocessing parameters leading to feature loss, single feature fusion dimension making it difficult to capture complex patterns, and fragmented model modules that cannot be co-optimized.
[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:
[0007] This invention relates to a neural stress injury recognition algorithm based on multi-attention crossover, comprising the following steps:
[0008] Acquire raw electroencephalogram (EEG) signal data;
[0009] Learnable bandpass filter parameters are introduced into the raw signal of each channel. These learnable bandpass filter parameters are represented in the form of nn.Parameter in the deep learning framework. During forward propagation, they are mapped to the [0.5, 40] Hz range through the Sigmoid function, and the low cutoff frequency and high cutoff frequency of the filter are determined based on this range.
[0010] A fourth-order Butterworth filter is used to perform bandpass filtering on the signals of each channel.
[0011] Empirical Mode Decomposition (EMD) is performed on the filtered signal to obtain multiple Intrinsic Mode Function (IMF) components. The energy value of each IMF component is calculated, and the IMF component with the highest energy is selected as the effective signal component to be retained.
[0012] The noise reduction threshold is dynamically adjusted based on the high-frequency energy of the signal. The threshold is calculated by combining a learnable initial threshold factor, high-frequency energy, maximum power spectrum, and a small constant to prevent division by zero. Threshold filtering is applied to the signal to suppress electromyographic artifacts.
[0013] The denoised signal is z-score normalized to calculate the mean and standard deviation of each channel signal, and the original signal is normalized. The normalized signal is then organized into a three-dimensional tensor with the shape of batch size, number of channels and number of sampling points.
[0014] Construct attention branches for time, space, and frequency respectively;
[0015] The temporal attention branch takes the mean of the input tensor along the channel dimension, generates temporal attention weights through a fully connected layer and a Sigmoid activation function, and multiplies them element-wise with the original tensor to obtain temporally weighted features.
[0016] The spatial attention branch takes the mean of the input tensor in the time dimension, generates spatial attention weights through a fully connected layer and a Sigmoid activation function, and multiplies them element-wise with the original tensor to obtain spatially weighted features;
[0017] The frequency attention branch performs a Fast Fourier Transform (FFT) on the input tensor to obtain the amplitude spectrum. After averaging the amplitude spectrum in the channel dimension, it generates frequency attention weights through a fully connected layer and a Sigmoid activation function. The weights are then transformed back to the time domain by an inverse FFT and multiplied element-wise with the original tensor to obtain the frequency-weighted features.
[0018] The time-weighted features, spatial-weighted features, and frequency-weighted features are input into the cross-gated fusion module. The learnable gating parameters are normalized by the Softmax function to generate three normalized fusion weights. The three features are then weighted and summed to output the fused damage-sensitive feature tensor.
[0019] The fused features are flattened into a one-dimensional vector and input into a spiking neural network consisting of two fully connected layers. Each layer contains multiple LIF neurons. The membrane potential update formula of the neurons includes the membrane potential at the previous time step, the input weighted sum, a learnable membrane potential decay coefficient, and a pulse feedback term. The firing threshold of the neurons is a learnable parameter used to generate pulse output.
[0020] The final output layer activates the pulse signal with Sigmoid to generate a probability output of neural stress injury; at the same time, the heatmap generation layer outputs a brain region localization heatmap and calculates the severity score of the injury.
[0021] The heatmap generation layer is a fully connected layer or a convolutional layer, used to generate a spatial distribution reflecting the contribution of electrodes in each brain region to damage discrimination from the output of the first layer of spiking neurons.
[0022] The learnable parameters in all modules are jointly optimized end-to-end using the backpropagation algorithm within a unified loss function framework. The loss function includes the binary cross-entropy main loss, the spatial consistency loss based on the mean of the first-layer impulse output in the batch dimension, and the L2 regularization term. The Adam or SGD optimizer is used to complete the parameter update.
[0023] The present invention is further configured such that the learnable bandpass filter parameters are calculated and updated through an automatic differentiation mechanism during backpropagation, thereby achieving adaptive learning of the frequency band.
[0024] The present invention is further configured such that the energy value of the IMF component is calculated by the sum of squares of the component signal, i.e. The IMF component with the highest energy value is selected as the effective signal component for subsequent processing.
[0025] The present invention is further configured such that the calculation formula for the dynamic threshold is:
[0026]
[0027] Where α is a learnable initial threshold factor, HighFreqEnergy represents the high-frequency energy, max(PowerSpectrum) is the maximum power spectrum value, and ∈ is a small constant to prevent division by zero.
[0028] The present invention is further configured such that the learnable gating parameters in the cross-gating fusion module are normalized by the Softmax function to generate a weight vector w = [w t ,w s ,w f ], satisfying w t +w s +w f=1, used for weighted fusion of time, space and frequency weighted features.
[0029] The present invention is further configured such that, after performing a fast Fourier transform, the frequency attention branch only uses the amplitude spectrum for processing and ignores the phase information, so as to enhance the robustness of the model to phase perturbations.
[0030] The present invention further specifies that the membrane potential update formula for each layer of the LIF neuron in the spiking neural network is:
[0031] V t =λV t-1 +W·x t -θ·o t-1
[0032] Among them, V t Let λ be the membrane potential at the current moment, λ be the learnable decay coefficient, W be the input weight, and x be the membrane potential. t As input, θ is a learnable firing threshold, o t-1 This is the pulse output from the previous moment.
[0033] The present invention is further configured such that the heat map output by the heat map generation layer is used to visualize the contribution of electrodes in different brain regions to damage identification, and to assist in clinical localization of damaged neural regions.
[0034] The present invention further specifies that the total loss function is defined as:
[0035]
[0036] in, For binary cross-entropy loss; Spatial consistency loss is defined as the minimization term of the feature variance of the first-layer pulse output in the batch dimension mean. β1, β2, and β3 are the L2 regularization terms for all learnable parameters; β1, β2, and β3 are preset weight coefficients.
[0037] The present invention is further configured such that, during the training process, the algorithm performs end-to-end joint updates of all learnable parameters in the adaptive preprocessing module, the multi-attention cross-fusion module, and the spiking neural network decision module through the backpropagation algorithm, with the optimizer being Adam and the learning rate ranging from [0.0001, 0.001].
[0038] The present invention has the following beneficial effects:
[0039] 1. This invention designs key preprocessing parameters such as the bandpass filter cutoff frequency, EMD component selection criteria, and denoising threshold factor as learnable and differentiable network parameters, and embeds them into an end-to-end training framework. These parameters are jointly updated with the main model parameters during backpropagation, and can be adaptively adjusted according to specific task objectives (such as damage identification), achieving a "tailor-made" enhancement of the original EEG signal. This mechanism effectively avoids feature loss caused by improper preprocessing in traditional processes, significantly improves the ability to capture weak and atypical neural stress signals, and enhances the feature fidelity and individual adaptability of the model.
[0040] 2. This invention proposes a three-way parallel attention mechanism encompassing time, space, and frequency, respectively capturing abnormal discharges in key time segments, the spatial distribution of abnormal brain regions, and energy changes in specific frequency bands (such as alpha inhibition and theta enhancement). By introducing a learnable cross-gated fusion module, the model can dynamically assess the importance of the three types of features and generate adaptive fusion weights. This mechanism not only achieves deep interaction and complementarity of multimodal information but also endows the model with stronger feature representation capabilities. Simultaneously, the visualization results of each attention weight can intuitively reflect the EEG features that the model focuses on, significantly improving the transparency and clinical interpretability of the model's decision-making process and providing doctors with reliable auxiliary diagnostic evidence.
[0041] 3. The LIF neurons in this invention possess learnable membrane potential decay coefficients and dynamic threshold parameters, enabling the network to adaptively adjust the "excitability" of neurons based on input data, greatly enhancing the model's expressive power and flexibility. Addressing the difficulty of SNN training, an innovative dual-mode gradient loss mechanism is proposed: temporal loss is based on backpropagation of membrane potential decay, ensuring the accuracy of temporal dynamic modeling; spatial loss minimizes the variance of the first-layer pulse output, encouraging spatial feature consistency and effectively alleviating the gradient vanishing problem caused by pulse sparsity. This mechanism significantly improves the convergence speed and stability of SNNs, achieving efficient training.
[0042] 4. This invention can simultaneously generate three key diagnostic information items: damage probability quantifies the individual risk level; damage localization heatmap accurately locates abnormal brain regions (such as the prefrontal and temporal lobes) by visualizing the spatial distribution of the first-layer impulse response, assisting doctors in lesion identification; severity score provides a continuous assessment of damage degree based on impulse firing rate or membrane potential accumulation. This multi-dimensional output mode comprehensively surpasses single classification results, providing richer and more refined decision support for clinical practice, and helping to achieve precision medicine and personalized intervention.
[0043] 5. This invention integrates three major modules—adaptive preprocessing, multi-attention fusion, and spiking neural network decision-making—into a unified differentiable framework, with all parameters jointly optimized under a unified loss objective. This end-to-end system design breaks through the limitations of traditional "module fragmentation," ensuring that information achieves co-evolution rather than isolated optimization during transmission. Signal preprocessing provides optimal input for feature fusion, feature fusion provides the most discriminative representation for SNN decision-making, and the feedback from the SNN, in turn, guides the adjustment of front-end parameters. This closed-loop optimization mechanism greatly improves the overall performance of the model, achieving a system-level leap in detection sensitivity, stability, generalization ability, and the ability to identify weak damage, providing a highly robust solution for the intelligent identification of neural stress and brain injury.
[0044] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.
[0046] Figure 1 This is a schematic diagram of the EEG signal preprocessing flow of the present invention.
[0047] Figure 2 This is a schematic diagram of the multi-attention cross-fusion of the present invention.
[0048] Figure 3 This is a schematic diagram of the decision-making process for neural stress injury recognition based on dynamic threshold spiking neurons according to the present invention.
[0049] Figure 4 This is a diagram illustrating the architecture of the learnable neural stress injury recognition system of the present invention. Detailed Implementation
[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
[0051] Example
[0052] Please see Figure 1-4 This invention is a neural stress injury recognition algorithm based on multi-attention crossover, comprising the following steps:
[0053] Acquire raw electroencephalogram (EEG) signal data;
[0054] Learnable bandpass filter parameters are introduced into the raw signal of each channel. These parameters are represented in the form of nn.Parameter in the deep learning framework and mapped to the [0.5, 40] Hz range via the Sigmoid function during forward propagation. The low and high cutoff frequencies of the filter are then determined based on this range.
[0055] f low =0.5 + 39.5·σ(θ) low ),f high =0.5 + 39.5·σ(θ) high )
[0056] in, f is the sigmoid function; low ,f high ∈[0.5,40]Hz, covering key EEG rhythms such as δ (0.5-4Hz), θ (4-8Hz), α (8-13Hz), β (13-30Hz), and low γ (30-40Hz);
[0057] A fourth-order Butterworth filter is used to perform bandpass filtering on the signals in each channel. The difference equation is as follows:
[0058]
[0059] in, This is the original signal from channel c; The filtered signal; a k b k By f low and f high Decide;
[0060] Empirical Mode Decomposition (EMD) is performed on the filtered signal to obtain multiple Intrinsic Mode Function (IMF) components. The energy value of each IMF component is calculated, and the IMF component with the highest energy is selected as the effective signal component. The energy value of each IMF component is calculated by the sum of the squares of its signal components. Select the IMF component with the highest energy value as the effective signal component for subsequent processing;
[0061] The denoising threshold is dynamically adjusted based on the high-frequency energy of the signal. The threshold is calculated from a learnable initial threshold factor, high-frequency energy, maximum power spectrum value, and a small constant to prevent division by zero. Threshold filtering is applied to the signal to suppress electromyographic artifacts. The formula for calculating the dynamic threshold is as follows:
[0062]
[0063] Where α is a learnable initial threshold factor, HighFreqEnergy represents the high-frequency energy, max(PowerSpectrum) is the maximum power spectrum value, and ∈ is a small constant to prevent division by zero;
[0064] The denoised signal is z-score normalized to calculate the mean and standard deviation of each channel signal, and the original signal is normalized. The normalized signal is then organized into a three-dimensional tensor with the shape of batch size, number of channels and number of sampling points.
[0065] Construct attention branches for time, space, and frequency respectively;
[0066] The temporal attention branch takes the mean of the input tensor along the channel dimension, generates temporal attention weights through a fully connected layer and a Sigmoid activation function, and multiplies them element-wise with the original tensor to obtain temporally weighted features.
[0067] The spatial attention branch takes the mean of the input tensor in the time dimension, generates spatial attention weights through a fully connected layer and a Sigmoid activation function, and multiplies them element-wise with the original tensor to obtain spatially weighted features;
[0068] The frequency attention branch performs a Fast Fourier Transform (FFT) on the input tensor to obtain the amplitude spectrum. After averaging the amplitude spectrum along the channel dimension, it generates frequency attention weights through a fully connected layer and a Sigmoid activation function. The weights are then transformed back to the time domain by an inverse FFT and multiplied element-wise with the original tensor to obtain the frequency-weighted features. After performing the Fast Fourier Transform, the frequency attention branch only uses the amplitude spectrum for processing and ignores phase information to enhance the robustness of the model to phase perturbations.
[0069] The time-weighted features, spatial-weighted features, and frequency-weighted features are input into the cross-gated fusion module. The learnable gating parameters are normalized by the Softmax function to generate three normalized fusion weights. The three features are then weighted and summed to output the fused damage-sensitive feature tensor.
[0070] The fused features are flattened into a one-dimensional vector and input into a spiking neural network consisting of two fully connected layers. Each layer contains multiple LeakyIntegrate-and-Fire (LIF) neurons. The membrane potential update formula of the neurons includes the membrane potential at the previous time step, the input weighted sum, a learnable membrane potential decay coefficient, and a pulse feedback term. The firing threshold of the neurons is a learnable parameter used to generate pulse output.
[0071] The final output layer applies Sigmoid activation to the pulse signal to generate a probability output of neural stress injury; simultaneously, a heatmap generation layer outputs a brain region localization heatmap and calculates the severity score of the injury. The learnable gating parameters in the cross-gated fusion module are normalized using the Softmax function to generate a weight vector w = [w t ,w s ,w f ], satisfying w t +w s +w f =1, used for weighted fusion of time, space, and frequency-weighted features; the update formula for the membrane potential of LIF neurons in each layer of the spiking neural network is:
[0072] V t =λV t-1 +W·x t -θ·o t-1
[0073] Among them, V t Let λ be the membrane potential at the current moment, λ be the learnable decay coefficient, W be the input weight, and x be the membrane potential. t As input, θ is a learnable firing threshold, o t-1 This is the pulse output from the previous moment;
[0074] The heatmap generation layer is a fully connected layer or a convolutional layer, used to generate a spatial distribution reflecting the contribution of electrodes in each brain region to damage discrimination from the output of the first layer of spiking neurons. The heatmap output by the heatmap generation layer is used to visualize the contribution of electrodes in different brain regions to damage identification and to assist in clinical localization of damaged nerve regions.
[0075] All learnable parameters in all modules are jointly optimized end-to-end using the backpropagation algorithm within a unified loss function framework. The loss function includes a binary cross-entropy main loss, a spatial consistency loss based on the batch dimension mean of the first-layer impulse output, and an L2 regularization term. Parameter updates are performed using either the Adam or SGD optimizer. The total loss function is defined as follows:
[0076]
[0077] in, For binary cross-entropy loss; Spatial consistency loss is defined as the minimization term of the feature variance of the first-layer pulse output in the batch dimension mean. β1, β2, and β3 are the L2 regularization terms for all learnable parameters; β1, β2, and β3 are preset weight coefficients.
[0078] Specifically, the learnable bandpass filter parameters are updated by calculating gradients through an automatic differentiation mechanism during backpropagation, thereby achieving adaptive learning of the frequency band. During training, the algorithm performs end-to-end joint updates of all learnable parameters in the adaptive preprocessing module, the multi-attention cross-fusion module, and the spiking neural network decision module through the backpropagation algorithm. The optimizer is Adam, and the learning rate range is [0.0001, 0.001].
[0079] The neural stress injury recognition algorithm based on multi-attention crossover provided in this embodiment constructs an end-to-end differentiable, system-level jointly optimized deep learning framework from raw EEG signal input to injury state output. By seamlessly integrating signal preprocessing, feature extraction, multimodal fusion, and brain-like decision-making mechanisms, it achieves high-sensitivity recognition of weak neural injury features. The algorithm first performs learnable adaptive preprocessing on the raw EEG signal: a differentiable bandpass filter module is introduced, and the learnable parameters are mapped to the [0.5, 40] Hz physiological frequency band using the Sigmoid function, followed by frequency domain filtering using a 4th-order Butterworth filter; then, Empirical Mode Decomposition (EMD) is performed on the filtered signal to calculate the energy of each intrinsic mode function (IMF) component and adaptively select the component with the highest energy as the effective signal; further, the noise reduction threshold is dynamically adjusted based on the high-frequency energy, where the threshold factor is a learnable parameter, and EMG artifact suppression is achieved by combining the maximum power spectrum value and a small constant to prevent division by zero; finally, the signal is z-score normalized to eliminate non-stationarity and output a structured tensor. In the feature fusion stage, the algorithm constructs a three-way attention mechanism in parallel: temporal, spatial, and frequency. Temporal attention generates temporal weights by applying a fully connected layer and sigmoid activation to the channel mean features, highlighting key time segments. Spatial attention generates spatial weights by performing a fully connected transformation on the temporal mean features, identifying abnormal brain region electrode distributions. Frequency attention performs a Fast Fourier Transform (FFT) on the signal, applies attention weights to the channel mean spectrum, and enhances the response in specific frequency bands after an inverse FFT transformation. The three weighted features are dynamically weighted through a cross-gated fusion module, where learnable gating parameters are normalized using Softmax to generate fusion weights, achieving adaptive integration of multi-dimensional features. The fused features are flattened and input into an innovatively designed hierarchical spiking neural network (SNN). This network uses a Leaky Integrate-and-Fire (LIF) neuron model, where the membrane potential decay coefficient and firing threshold are learnable parameters. A two-layer fully connected structure enables pulse coding and decision output. Finally, the network outputs the damage probability, a brain region localization heatmap generated based on the first layer's pulse response, and a damage severity score. The entire model is trained end-to-end using a unified joint loss function, which includes a binary cross-entropy main loss, a spatial consistency loss based on the impulse output mean, and an L2 regularization term. All learnable parameters in all modules (including filter parameters, threshold factors, attention weights, gating coefficients, SNN dynamic parameters, etc.) are jointly updated through an automatic differentiation mechanism during backpropagation. The Adam optimizer is used to optimize the parameters, thereby achieving the co-evolution of signal enhancement, feature fusion, and brain-like decision-making, significantly improving the accuracy and robustness of identifying neural stress and brain injury patterns.
[0080] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
Claims
1. A neural stress injury recognition algorithm based on multi-attention crossover, characterized in that, Includes the following steps: Acquire raw electroencephalogram (EEG) signal data; Learnable bandpass filter parameters are introduced into the raw signal of each channel. These learnable bandpass filter parameters are represented in the form of nn.Parameter in the deep learning framework. During forward propagation, they are mapped to the [0.5, 40] Hz range through the Sigmoid function, and the low cutoff frequency and high cutoff frequency of the filter are determined based on this range. A fourth-order Butterworth filter is used to perform bandpass filtering on the signals of each channel. Empirical Mode Decomposition (EMD) is performed on the filtered signal to obtain multiple Intrinsic Mode Function (IMF) components. The energy value of each IMF component is calculated, and the IMF component with the highest energy is selected as the effective signal component to be retained. The noise reduction threshold is dynamically adjusted based on the high-frequency energy of the signal. The threshold is calculated by combining a learnable initial threshold factor, high-frequency energy, maximum power spectrum, and a small constant to prevent division by zero. Threshold filtering is applied to the signal to suppress electromyographic artifacts. The denoised signal is z-score normalized to calculate the mean and standard deviation of each channel signal, and the original signal is normalized. The normalized signal is then organized into a three-dimensional tensor with the shape of batch size, number of channels and number of sampling points. Construct attention branches for time, space, and frequency respectively; The temporal attention branch takes the mean of the input tensor along the channel dimension, generates temporal attention weights through a fully connected layer and a Sigmoid activation function, and multiplies them element-wise with the original tensor to obtain temporally weighted features. The spatial attention branch takes the mean of the input tensor in the time dimension, generates spatial attention weights through a fully connected layer and a Sigmoid activation function, and multiplies them element-wise with the original tensor to obtain spatially weighted features; The frequency attention branch performs a Fast Fourier Transform (FFT) on the input tensor to obtain the amplitude spectrum. After averaging the amplitude spectrum in the channel dimension, it generates frequency attention weights through a fully connected layer and a Sigmoid activation function. The weights are then transformed back to the time domain by an inverse FFT and multiplied element-wise with the original tensor to obtain the frequency-weighted features. The time-weighted features, spatial-weighted features, and frequency-weighted features are input into the cross-gated fusion module. The learnable gating parameters are normalized by the Softmax function to generate three normalized fusion weights. The three features are then weighted and summed to output the fused damage-sensitive feature tensor. The fused features are flattened into a one-dimensional vector and input into a spiking neural network consisting of two fully connected layers. Each layer contains multiple LIF neurons. The membrane potential update formula of the neurons includes the membrane potential at the previous time step, the input weighted sum, a learnable membrane potential decay coefficient, and a pulse feedback term. The firing threshold of the neurons is a learnable parameter used to generate pulse output. The final output layer activates the pulse signal with Sigmoid to generate a probability output of neural stress injury; at the same time, the heatmap generation layer outputs a brain region localization heatmap and calculates the severity score of the injury. The heatmap generation layer is a fully connected layer or a convolutional layer, used to generate a spatial distribution reflecting the contribution of electrodes in each brain region to damage discrimination from the output of the first layer of spiking neurons. The learnable parameters in all modules are jointly optimized end-to-end using the backpropagation algorithm within a unified loss function framework. The loss function includes the binary cross-entropy main loss, the spatial consistency loss based on the mean of the first-layer impulse output in the batch dimension, and the L2 regularization term. The Adam or SGD optimizer is used to complete the parameter update.
2. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The learnable bandpass filter parameters are updated during backpropagation by calculating gradients through an automatic differentiation mechanism, thereby achieving adaptive learning of the frequency band.
3. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The energy value of the IMF component is calculated by the sum of the squares of the component signal, i.e. The IMF component with the highest energy value is selected as the effective signal component for subsequent processing.
4. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The formula for calculating the dynamic threshold is: Where α is a learnable initial threshold factor, HighFreqEnergy represents the high-frequency energy, max(PowerSpectrum) is the maximum power spectrum value, and ∈ is a small constant to prevent division by zero.
5. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The learnable gating parameters in the cross-gating fusion module are normalized using the Softmax function to generate a weight vector w = [w t ,w s ,w f ], satisfying w t +w s +w f =1, used for weighted fusion of time, space and frequency weighted features.
6. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The frequency attention branch processes only the amplitude spectrum after performing the fast Fourier transform, ignoring phase information, in order to enhance the model's robustness to phase perturbations.
7. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The formula for updating the membrane potential of LIF neurons in each layer of the spiking neural network is as follows: V t =λV t-1 +W x t -θ o t-1 Among them, V t Let λ be the membrane potential at the current moment, λ be the learnable decay coefficient, W be the input weight, and x be the membrane potential. t As input, θ is a learnable firing threshold, o t-1 This is the pulse output from the previous moment.
8. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The heatmaps output by the heatmap generation layer are used to visualize the contribution of electrodes in different brain regions to damage identification, and to assist in clinical localization of damaged neural regions.
9. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, The total loss function is defined as follows: in, For binary cross-entropy loss; Spatial consistency loss is defined as the minimization term of the feature variance of the first-layer pulse output in the batch dimension mean. β1, β2, and β3 are the L2 regularization terms for all learnable parameters; β1, β2, and β3 are preset weight coefficients.
10. The neural stress injury recognition algorithm based on multi-attention crossover according to claim 1, characterized in that, During training, the algorithm performs end-to-end joint updates of all learnable parameters in the adaptive preprocessing module, the multi-attention cross-fusion module, and the spiking neural network decision module using the backpropagation algorithm. The optimizer is Adam, and the learning rate ranges from [0.0001, 0.001].