Acupuncture brain machine interface system based on brain-like recursive network model decoding

Abstract

The invention relates to an acupuncture brain machine interface system based on brain-like recursive network model decoding. The system comprises an electroencephalogram acquisition module; an electroencephalogram preprocessing module; the needling parameter fitting module is used for accurately fitting needle body motion parameters through a polynomial regression method; the state space model is used for accurately simulating the nonlinear change characteristic of the neuron membrane potential according to the needle body motion parameters output by the acupuncture parameter fitting module; the brain-like recursive network model is used for depicting time dependence of neural activities of a single brain region under the acupuncture action and correlation of neural activities among different brain regions; a Kalman filtering generator; and the neural decoder is used for performing behavior decoding on the optimal internal state output by the brain-like recursive network model updated by the Kalman filtering generator. The system integrates electroencephalogram signal acquisition, acupuncture parameter acquisition, brain nerve activity modeling and acupuncture manipulation decoding, can perform real-time online operation, and is high in time resolution and high in response speed.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface technology, specifically to an acupuncture brain-computer interface system based on a brain-like recursive network coding model. Background Technology

[0002] In the field of Traditional Chinese Medicine (TCM) acupuncture, needling is the core component of the acupuncture process. The key to needling lies in the acupuncturist's precise control of the filiform needles through finger movements, achieving the interaction between the needles and acupoints to ultimately achieve therapeutic goals. However, currently, the efficacy and quantification of acupuncture techniques rely heavily on the subjective feelings of both the acupuncturist and the patient. This places high demands on the acupuncturist's experience and skills and is significantly influenced by individual patient differences, severely hindering the intelligent development of acupuncture.

[0003] Brain-computer interface (BCI) refers to a direct connection created between the brain and an external device, capable of extracting signals from the brain and converting them into recognizable commands, enabling information exchange between the brain and the device. Currently, BCI technology shows great application potential in many fields such as entertainment, education, and medicine, and non-invasive BCI based on electroencephalography (EEG) has received widespread attention due to its high safety and convenience. By preprocessing the acquired EEG signals and extracting key features, combined with advanced decoding algorithms, it is possible to accurately interpret brain intentions from EEG signals, achieving precise evaluation of acupuncture efficacy and accurate quantification of acupuncture techniques. However, due to the influence of scalp obstruction and electromagnetic interference, the effective features in the acquired EEG data are severely weakened, and the EEG manifestations of the same behavior vary greatly among different individuals. Even for the same individual, signals can change at different times and under different states. These challenges pose significant difficulties for the development of high-precision decoding algorithms.

[0004] The core of BCI decoding is establishing the connection between neural activity and external behavior, and uncovering deeper patterns of neural activity is key to improving decoding accuracy. EEG, in essence, is a macroscopic representation of mesoscopic and microscopic neural activity within the brain. If we can further obtain patterns of neural activity within the brain based on EEG signals, we can effectively eliminate the interference of irrelevant noise such as electromyography and electrooculography on core neural signals, establishing a more precise causal mapping between neural activity and motor intention, thereby improving the robustness and accuracy of decoding. Previous research has attempted to simulate cortical neural activity by establishing biological mechanism models, resulting in a series of neural models such as the Hodgkin-Huxley model and the Neural Mass model. These models accurately characterize the dynamic mechanisms of neuronal ion channels, large-scale collective neural behavior, and other levels of neural activity. However, such models often face bottlenecks in training and learning. How to make the model closely resemble brain-like mechanisms while maintaining strong learning capabilities remains an urgent problem to be solved.

[0005] With the increasing maturity of deep learning, current research focuses on developing powerful deep neural networks to learn the patterns of neural activity. Deep neural networks, with their strong nonlinear modeling and high-dimensional parameter processing capabilities, exhibit greater adaptability and robustness in the face of individual differences and dynamic noise. In recent years, research teams both domestically and internationally have proposed neural dynamic modeling methods based on convolutional neural networks, recurrent neural networks, Transformers, and generative adversarial networks. However, most of these network model architectures follow traditional layouts, using multi-layer convolutions and activation functions to achieve feature extraction and nonlinear approximation, failing to truly form a brain-like working paradigm. This keeps the models at the level of feature learning, unable to truly understand the mechanisms behind neural activity, thus making it difficult to function like the brain. Furthermore, facing the complex and ever-changing brain activity, if the model only learns its features rather than its mechanisms, it will struggle to handle the flexible scenarios that may arise in brain-computer interface applications, reducing the generalization and practicality of brain-computer interface systems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention aims to provide an acupuncture brain-computer interface system based on a brain-like recurrent network model decoding. This system achieves the conversion of acupuncture parameters into afferent neuron membrane potential characteristics through a state-space model. Simultaneously, it uses a brain-like recurrent network model as its core to accurately simulate the interaction mechanism between excitatory and inhibitory neurons in multiple brain regions. The brain-like recurrent network model architecture integrates biological constraints such as brain-like connectivity constraints and excitation-inhibition balance, making the model more brain-like and improving the accuracy of brain neural activity modeling using EEG data. This lays a solid foundation for accurate decoding of non-invasive acupuncture brain-computer interfaces based on EEG. The system integrates EEG signal acquisition, acupuncture parameter acquisition, brain neural activity modeling, and acupuncture technique decoding, enabling real-time online operation with high temporal resolution and fast response speed.

[0007] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows: This invention provides an acupuncture brain-computer interface system based on a brain-like recurrent network model, the system comprising: The EEG acquisition module is used to acquire raw EEG signals; The EEG preprocessing module is used to acquire clean EEG signals. ; The acupuncture parameter fitting module includes acupuncture parameters such as rotation angle. Stimulation frequency Stimulation depth The needle motion parameters were accurately fitted using the polynomial regression method. The state-space model, which includes a linear time-varying state equation and a nonlinear readout part, is used to fit the needle motion parameters output by the acupuncture parameter fitting module. It captures the linear characteristics of needle motion through the linear time-varying state-space equation and uses the tanh function and ReLU function to construct a piecewise activation function to read out the internal linear state nonlinearly, accurately simulating the nonlinear change characteristics of neuronal membrane potential. The brain-like recursive network model is constructed based on the connectivity map between different brain regions of the cerebral cortex and the coupling relationship between excitatory and inhibitory neurons in the cortex. It consists of several brain regions and can characterize the time dependence of neural activity in a single brain region and the correlation between neural activities in different brain regions under acupuncture. After constraining the output to make it close to reality, the output is then used to fit the acupuncture parameters.

[0008] The Kalman filter generator consists of alternating prediction generation and update iteration phases. In the prediction generation phase, the initial state distribution is simulated using the output of the neuromorphic recurrent network model, and the model is run using the current parameters of the neuromorphic recurrent network model. a time step until t At that moment, what you have gained The combination of internal state outputs yields the steady-state activity distribution of excitatory and inhibitory neurons in the brain, updating the internal network parameters of the brain-like recurrent network model; The neural decoder is used to perform behavioral decoding on the optimal internal state output by the brain-like recurrent network model updated by the Kalman filter generator, and output the predicted needle-prick parameters.

[0009] The brain-like recurrent network model includes an excitation-inhibition dual recurrent neural network for different brain regions. The excitation-inhibition dual recurrent neural network includes an excitatory recurrent network (E-RNN) and an inhibitory recurrent network (I-RNN). Both E-RNN and I-RNN use 128 hidden state units to describe the state changes of excitatory neurons and inhibitory neurons themselves. There are interactive connections between E-RNN and I-RNN to characterize the interaction between excitatory neurons and inhibitory neurons. With each brain region in t The quotient of the excitatory and inhibitory hidden states in a time-varying excitation-inhibition dual recurrent neural network is used as the average outward network output of that brain region. ; Setting excitation-inhibition balance threshold constraints for each brain region , Average output of the network outward from each brain region By maintaining dimensional consistency, the network's internal parameters are optimized in reverse based on the residuals between the two, so that the average output of the network in a single brain region maintains an excitatory-inhibitory balance. Brain-like connectivity constraints that determine the interconnections between excitation-inhibition dual recurrent neural networks in different brain regions using cortical connectivity maps constraining brain-like connectivity The internal state output of the entire brain-like recurrent network model at time t is obtained by matrix multiplication with the average output of all brain regions that satisfy the excitation-inhibition balance constraint.

[0010] The neural decoder comprises an input dimensionality reduction module, a temporal feature extraction module, and an output probability module. The input dimensionality reduction module uses a single-layer fully connected structure to process the optimized neural network model... t Internal state at any moment Dimensionality reduction mapping is performed, and the internal state after dimensionality reduction is used as the initial hidden state. The temporal feature extraction module uses a 3-layer gated recurrent unit (GRU) to extract temporal dependent features step by step and outputs three needle-prick parameters (used for decoding to obtain the predicted needle-prick parameters).

[0011] Compared with the prior art, the beneficial effects of the present invention are: (1) The method of this invention changes the traditional approach of relying on deep convolutional loops to extract effective features in deep learning modeling. It designs a deep learning network framework based on brain-like mechanisms, which can more accurately learn the changing patterns of neural activity inside the brain. Faced with the conversion from acupuncture parameters to neuronal membrane potential features, the constructed state-space model adopts a fusion mode of linear part extraction of simple features and nonlinear readout reconstruction of complex details. This fully preserves the effective features and avoids overfitting due to excessive model complexity.

[0012] (2) The brain-like recursive network model constructed in this invention takes into account the characteristics of brain activity, namely, there are two major types of neurons in brain activity: excitatory neurons and inhibitory neurons. These two types of neurons are distributed in various brain regions. If we want to accurately simulate brain activity, we can consider simulating the recursive and interaction mechanism between these two types of neurons to realize the activity of a certain brain region. Then, we can realize the reproduction of the common activity of multiple brain regions through brain-like connection constraints. Two RNN networks are constructed for a single brain region. One RNN network simulates the activity of excitatory neurons and the other RNN network simulates the activity of inhibitory neurons. There are also interactive connections between the two RNN networks to simulate the interaction between the two major types of neurons. At the same time, excitatory-inhibitory balance constraints are added to regulate network activity. After the network of a single brain region is constructed, the connection between multiple brain region networks is realized by adding brain-like connection constraints, and finally the accurate simulation of whole brain activity is achieved. Compared with other brain-like models, the brain-like recurrent network model of this invention focuses on the activity characteristics between two major neuron clusters and incorporates excitation-inhibition balance constraints and brain-like connectivity constraints. It adopts a dual RNN network structure and integrates two major biological constraints: brain-like connectivity constraints and excitation-inhibition balance constraints. This can regulate the state update mode of the network, making the network model more closely resemble the working mechanism of the real brain.

[0013] (3) The brain-like recurrent network model of this invention fully considers the excitation-inhibition balance in neural activity, using dual RNN networks to model excitatory neurons and inhibitory neurons respectively, and then achieving organic connections between 210 brain region network models through brain-like connectivity constraints. Compared with traditional mechanistic models, the brain-like recurrent network model based on RNN networks has a strong advantage in processing temporal information and time dependence, and can better simulate the complex interaction between excitatory and inhibitory neurons. In addition, compared with brain-computer interface systems that directly decode motor behavior based on EEG signals, this invention achieves direct decoding of motor behavior from the internal state of the brain by deeply restoring the neural activity inside the brain. This breaks through the bottleneck of incorrect decoding due to insufficient accuracy of EEG signals and provides a new approach for BCI decoding.

[0014] (4) The present invention uses a Kalman filter to realize the internal state update of the brain-like recurrent network model. Compared with the traditional method of using the loss function to drive the backpropagation of error to update the RNN network parameters, the state update based on the Kalman filter has better stability, and stability is crucial for BCI decoding.

[0015] (5) The core advantage of this invention lies in its ability to achieve real-time online operation integrating EEG signal acquisition, acupuncture parameter acquisition, brain neural activity modeling, and behavior decoding. Simultaneously, the trained brain-like recurrent network model can generate realistic internal brain neural activity under unsupervised conditions, providing a continuous stream of rich data for BCI decoding. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall framework of the brain-computer interface system of the present invention.

[0017] Figure 2 This is a detailed diagram of a brain-like recurrent network model. Detailed Implementation

[0018] The present invention will be further explained below with reference to the embodiments and accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0019] Example 1: This invention relates to an acupuncture brain-computer interface system based on a brain-like recurrent network model (see...). Figure 1-2 The system includes: The EEG acquisition module is used to acquire raw EEG signals; The EEG preprocessing module is used to acquire clean EEG signals. ; The acupuncture parameter fitting module includes acupuncture parameters such as rotation angle. Stimulation frequency Stimulation depth The needle motion parameters are accurately fitted using polynomial regression and denoted as the input vector. Polynomial regression represents the relationship between acupuncture parameters and needle motion parameters, as shown below: (1) in, , , All are learnable weight parameters; The state-space model comprises a linear time-varying state equation and a nonlinear readout component. It captures the linear characteristics of needle motion through the linear time-varying state-space equation and uses a piecewise activation function composed of tanh and ReLU functions to perform nonlinear readout of the internal linear state, accurately simulating the nonlinear changes in neuronal membrane potential. The state-space model uses the fitted needle motion parameters as an input vector, specifically represented as follows: (2) (3) (4) (5) in, Representing the state-space model in t The internal states at any given time include the Merkel cell membrane potential in the epidermis, the Schwann cell membrane potential, and the dorsal root ganglion PV. +Neuronal membrane potential and spinal cord X + The neuronal membrane potential has four state components. The state update matrix is ​​used to describe the autonomous evolution of the internal state. The matrix adopts a semi-parametric design that combines fixed physiological parameters with learnable adjustable parameters. The elements on the main diagonal of the matrix are quantitatively obtained based on a large number of experimental patterns and serve as fixed physiological parameters, while the remaining elements are learnable parameters. For the input transformation matrix, elements , , and All of these are learnable parameters. Representing the state-space model in t+ The internal state at time 1. This represents a piecewise activation function. x Represents a variable. This indicates a nonlinear output, i.e., the output of the state-space model.

[0020] In this invention, the needle motion can be viewed as a waveform, containing both nonlinear characteristics (such as curved portions) and linear characteristics (approximately straight portions). The linear time-varying state-space equation primarily captures the linear characteristics of the needle motion, i.e., using combinations of straight lines to replace curves as much as possible. By using piecewise activation functions instead of a single activation function, the action potentials of neurons can be simulated more accurately.

[0021] A brain-like recurrent network model was constructed based on the connectivity map between different brain regions of the cerebral cortex and the coupling relationships between excitatory and inhibitory neurons in the cortex. This model, consisting of 210 brain regions, characterizes the time dependence of neural activity in a single brain region under acupuncture and the correlation of neural activity between different brain regions. For a specific brain region, leveraging the excellent memory and powerful temporal processing capabilities of recurrent neural networks (RNNs), an excitatory recurrent network (E-RNN) and an inhibitory recurrent network (I-RNN) were constructed, forming a dual-coupled recurrent neural network for that single brain region—an excitatory-inhibitory dual recurrent neural network. Both the E-RNN and I-RNN employ 128 hidden state units to describe the state changes of excitatory and inhibitory neurons. Interactive connections exist between the E-RNN and I-RNN to characterize the interactions between excitatory and inhibitory neurons. m The state update process of the excitation-inhibition dual recurrent neural network in each brain region can be specifically represented as follows: (6) (7) (8) in, and They represent the first m Each brain region in t The excitatory and inhibitory hidden states of the 128 hidden state units at time 1; and They represent the first m Each brain region in t+ The excitatory and inhibitory hidden states of the 128 hidden state units at time 1; and They represent the first m Each brain region in t The excitatory hidden state and inhibitory hidden state of the first hidden state unit at time 1; represent t The excitatory self-connection matrix at time t; represent t The inhibitory self-connection matrix at time step; represent t The excitation-inhibition cross-connection matrix at any given time; and Represent t The excitatory control matrix and inhibitory control matrix at each moment; and The first m Excitatory and inhibitory neurons in each brain region t Random neural noise generated at any given moment; Take the first m Each brain region in t The quotient of the excitatory and inhibitory hidden states in a time-varying excitation-inhibition dual recurrent neural network is used as the average outward network output of that brain region. , denoted as: (9) in, This indicates that the elements at each corresponding position in the matrix are multiplied.

[0022] Setting excitation-inhibition balance threshold constraints for each brain region , and Maintaining dimensionality consistency, the network's internal parameters are inversely optimized based on the residuals between the two, ensuring that the average network output of a single brain region maintains an excitation-inhibition balance. Brain-inspired connectivity constraints are applied to different network connections in 210 brain regions throughout the brain. , The dimension of the matrix is The values ​​of the elements within the matrix are determined by the cerebral cortex connectivity map, representing the connection strength between different brain regions. The entire brain-like recurrent network model is then... t The internal state output at time t is denoted as Specifically, it can be represented as follows: (10) The Kalman filter generator consists of alternating prediction generation and update iteration phases. In the prediction generation phase, an initial state distribution is generated using a brain-like recurrent network model, and the model is run with the current parameters of the brain-like recurrent network model. a time step until t At that moment, what you have gained The combination of internal state outputs yields the steady-state activity distribution of excitatory and inhibitory neurons in the brain. Assuming t The activity status determined by real EEG at any given moment The current steady-state activity distribution is satisfied, that is: ,in The covariance matrix, determined by the steady-state activity distribution, is used to measure how close the estimated value is to the true value.

[0023] Current steady-state activity distribution Propagation to the prior distribution of the next time step The calculation process is as follows: (11) (12) in, Indicates the calculation of covariance; for t+ The activity state determined by the actual EEG at any given moment; Represent the prior covariance matrix; It expresses expectation.

[0024] During the update iteration phase, new measurements are used to correct prior predictions. This phase includes calculating the Kalman prediction error and adjusting the state estimate according to the proportion of this error.

[0025] Given prior prediction and new measurement EEG signals Prediction error It can be represented as: (13) in, This is the observation matrix.

[0026] By this error Multiply by the Kalman gain matrix To perform Kalman correction and form an updated state estimate : (14) The Kalman gain matrix is ​​determined using formula (15). : (15) in To control The weight matrix.

[0027] By using formulas (13)-(15), Minimize, achieve minimizing the correction estimate Uncertainty in the process.

[0028] Optimize at this time This problem constitutes least squares regression, and therefore can be solved using the following expression: (16) (17) in, This represents the random noise added during the estimation process.

[0029] Sure Then, the measured values ​​were included. Perform posterior estimation to obtain the posterior estimated covariance matrix. The calculation is as follows: (18) (19) in, Represents the identity matrix.

[0030] After this correction, the Kalman filter continues for one time step. The predictions, measurements, and corrections are evaluated. A Kalman filter is used to update the network parameters of the brain-like recurrent network model's internal state, thereby obtaining the optimal internal state for behavior decoding.

[0031] The neural decoder comprises an input dimensionality reduction module, a temporal feature extraction module, and an output probability module. The input dimensionality reduction module uses a single-layer fully connected structure to process the optimized features from the brain-like recurrent network model. t Internal state at any moment Perform dimensionality reduction mapping, and use the reduced internal state as the initial hidden state. The temporal feature extraction module employs a three-layer gated recurrent unit (GRU) to progressively extract temporal dependency features. The first GRU layer has a hidden dimension of 64, the second layer has 32, the third layer has 16, and so on. l ( l The calculation process of the GRU layer (=1, 2, 3) can be represented as follows: (20) (twenty one) (twenty two) (twenty three) in, This indicates a reset option, controlling the percentage of historical hidden states that are retained. This represents the update gate, which controls the fusion ratio between the current input and the historical hidden state; Indicates the candidate hidden state; Output the final hidden state; and To reset the weight matrix of the gate; To reset the door's bias; Use the Sigmoid activation function; For the first l layer The hidden state at any given moment; and To update the weight matrix of the gate; To update the gate's bias term; and This is the weight matrix for the candidate states; The bias term for the candidate state; This indicates that the elements at corresponding positions in the matrix are multiplied. The hyperbolic tangent activation function is used for the output of the last GRU layer. Hidden state at each time step The compressed temporal feature vector is obtained by averaging. The expression is as follows: (twenty four) The output probability module contains two layers of non-linear fully connected mappings, receiving... The input is three acupuncture parameters, which are calculated as follows: (25) (26) in, This is the weight matrix; For bias; For activation functions; , and This represents three acupuncture parameters (intensity, frequency, and depth) in t The predicted value at any given time; , and This is the weight vector for each parameter branch; , and This is a bias term.

[0032] Define loss function We employ a weighted mean square error to ensure a balance in the decoding accuracy of the three needle-punching parameters. (27) in, This represents the total number of time steps.

[0033] After calculating the loss, the internal parameters of the neural decoder are optimized using a backpropagation strategy to achieve accurate decoding.

[0034] In the training phase of the brain-like recurrent network model, this invention requires the use of EEG signals acquired by the EEG acquisition module, which are then preprocessed and used as real signal labels. After training, the EEG signals can be omitted, and the trained brain-like recurrent network model can be directly used to input the acupuncture parameters acquired in real time during acupuncture. After fitting and state space model processing, the parameters are fed into the brain-like recurrent network model. Finally, the decoder obtains the predicted acupuncture parameters, which can be used to correct the current acupuncture parameters online in real time, thereby improving the accuracy of acupuncture control.

[0035] The input of the brain-like recurrent network model of the present invention is derived from the fitting of acupuncture parameters, rather than electroencephalogram (EEG). Once the brain-like recurrent network model is trained to maturity, it can autonomously simulate brain activity to achieve behavioral decoding without relying on additional EEG recordings to drive it, thus having greater practicality.

[0036] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A needle-based brain-computer interface system based on brain-like recursive network model decoding, characterized in that, The system includes: The EEG acquisition module is used to acquire raw EEG signals; The EEG preprocessing module is used to acquire clean EEG signals. ; The acupuncture parameter fitting module includes acupuncture parameters such as rotation angle. Stimulation frequency and the depth of stimulation The needle motion parameters were accurately fitted using the polynomial regression method. The state-space model is used to fit the needle motion parameters output by the acupuncture parameter fitting module. It captures the linear characteristics of needle motion through a linear time-varying state-space equation and uses the tanh function and ReLU function to construct a piecewise activation function to read out the internal linear state nonlinearly, accurately simulating the nonlinear change characteristics of neuronal membrane potential. The brain-like recursive network model is constructed based on the connectivity map between different brain regions of the cerebral cortex and the coupling relationship between excitatory and inhibitory neurons in the cortex. It consists of several brain regions and can characterize the time dependence of neural activity in a single brain region and the correlation between neural activity in different brain regions under the action of acupuncture. A Kalman filter generator is used to simulate and generate an initial state distribution using the output of a neuromorphic recurrent network model, and then runs the model with the current parameters of the neuromorphic recurrent network model. a time step until t At that moment, what you have gained The combination of internal state outputs yields the steady-state activity distribution of excitatory and inhibitory neurons in the brain, updating the network parameters of the internal state of the brain-like recurrent network model. The neural decoder is used to perform behavioral decoding on the optimal internal state output by the brain-like recurrent network model updated by the Kalman filter generator, and output the predicted needle-prick parameters.

2. The system according to claim 1, characterized in that, The brain-like recurrent network model includes an excitation-inhibition dual recurrent neural network for different brain regions. The excitation-inhibition dual recurrent neural network includes an excitatory recurrent network (E-RNN) and an inhibitory recurrent network (I-RNN). Both E-RNN and I-RNN use 128 hidden state units to describe the state changes of excitatory neurons and inhibitory neurons themselves. There are interactive connections between E-RNN and I-RNN to characterize the interaction between excitatory neurons and inhibitory neurons. With each brain region in t The quotient of the excitatory and inhibitory hidden states in a time-varying excitation-inhibition dual recurrent neural network is used as the average outward network output of that brain region. ; Setting excitation-inhibition balance threshold constraints for each brain region , Average output of the network outward from each brain region By maintaining dimensional consistency, the network's internal parameters are optimized in reverse based on the residuals between the two, so that the average output of the network in a single brain region maintains an excitatory-inhibitory balance. Brain-like connectivity constraints that determine the interconnections between excitation-inhibition dual recurrent neural networks in different brain regions using cortical connectivity maps constraining brain-like connectivity The internal state output of the entire brain-like recurrent network model at time t is obtained by matrix multiplication with the average output of all brain regions that satisfy the excitation-inhibition balance constraint.

3. The system according to claim 1, characterized in that, The neural decoder comprises an input dimensionality reduction module, a temporal feature extraction module, and an output probability module; the input dimensionality reduction module uses a single fully connected layer to process the optimized neural network model... t Internal state at any moment Perform dimensionality reduction mapping, and use the reduced internal state as the initial hidden state; The temporal feature extraction module uses a 3-layer gated recurrent unit (GRU) to extract temporal dependent features step by step and outputs three needle-punching parameters.

4. The system according to claim 1, characterized in that, In the acupuncture parameter fitting module, the needle motion parameters are accurately fitted using polynomial regression and denoted as the input vector. Multinomial regression is specifically expressed as: (1) in, , , All of these are learnable weight parameters.

5. The system according to claim 1, characterized in that, The processing steps for the state-space model are formulas (2) and (5). (2) (5) in, Representing the state-space model in t The internal states at any given time include the Merkel cell membrane potential in the epidermis, the Schwann cell membrane potential, and the dorsal root ganglion PV. + Neuronal membrane potential and spinal cord X + The four state components of neuronal membrane potential; This is the state update matrix, used to describe the autonomous evolution of the internal states; is the input transformation matrix, and are the learnable parameters; This represents a piecewise activation function. x Represents variables; This represents the output of the state-space model.

6. The system according to claim 1, characterized in that, The loss function of the decoder for: (25) in, ^ represents the total number of time steps; ^ represents the predicted value of the corresponding acupuncture parameter.

Images