Dexterous hand motion decoding method based on collaborative base manifold

Abstract

The invention discloses a dexterous hand motion decoding method based on a collaborative base manifold, which comprises the following steps: acquiring electroencephalogram signal characteristics and corresponding hand motion data, and performing preprocessing after time sequence alignment to obtain a training data set; a dexterous hand motion decoding model is constructed, and in the model, a periodic network module maps the electroencephalogram signal features to a phase manifold to obtain electroencephalogram features with periodic features; the control network module extracts decoupled space-time motion control parameters; the collaborative base recruitment module carries out decoding according to the space-time motion control parameters to obtain corresponding hand motion; constructing total loss, and training the dexterous hand motion decoding model by using the training data set; and extracting electroencephalogram signal characteristics to be predicted, inputting the electroencephalogram signal characteristics to the trained dexterous hand motion decoding model, and decoding to obtain corresponding hand motion. According to the invention, more accurate and fine hand motion decoding can be realized.

Description

Technical Field

[0001] The invention belongs to the field of motor nerve signal decoding, and in particular relates to a dexterous hand movement decoding method based on collaborative basis manifolds. Background Art

[0002] Invasive brain-computer interfaces build direct communication and control pathways between the brain and external devices, and have shown important application value in medical fields such as motor function rehabilitation. As the core carrier of human interaction with the outside world, human hands have the ability to manipulate objects with high flexibility, precision, and almost no deliberate control. This fine hand movement control function plays an irreplaceable role in daily life. Given the fundamental position of hand function in human behavior, how to achieve continuous and precise reconstruction of hand movement function has become a key technical challenge to help disabled people restore their independent living and professional abilities. At present, this field has become a key frontier cross-cutting direction of global research.

[0003] For example, the Chinese patent document with publication number CN113101021A discloses a robotic arm control method based on the MI-SSVEP hybrid brain-computer interface, which realizes the control of the robotic arm through non-invasive EEG data acquisition, online EEG signal preprocessing, online EEG signal processing and decoding, generating user intended targets and converting the targets into control instructions.

[0004] Current brain-computer interface research has progressed to the stage of initially achieving dexterous control of robotic arms, that is, high-performance control of 3-4 degrees of freedom has been achieved, and autonomous coffee drinking and eating have been achieved. However, the human hand has 27 degrees of freedom, and dexterous control of the hand is extremely difficult. The main challenge lies in the coordination between multiple joints of the hand and the nonlinear interaction. Most of the current hand motion decoding work regards the parameters of each joint as an independent variable, ignoring the synergistic relationship between joints, and thus has limited effect on hand motion decoding. For this reason, some studies have proposed that the brain may regulate high-degree-of-freedom motion in a low-dimensional control space (manifold) and mine low-dimensional representations from neural signals. Although many different manifolds have been found in neural signals, the correspondence between these manifolds and hand motion control parameters is still unclear.

[0005] On the other hand, although hand motion control has high-dimensional degrees of freedom, these degrees of freedom are not completely independent, but redundant. Therefore, some works are based on dimensionality reduction methods to identify low-dimensional motion manifolds from complex hand movements, and decode high-dimensional motion based on this. However, since the learning process is highly dependent on assumptions and constraints, different motion manifolds will be obtained under different assumptions, which is obviously inconsistent with the brain's operational control mechanism. In addition, most studies only consider the motion signal itself and do not associate it with brain signals, so it is not yet known whether and how the motion manifold and neural signals are related to each other.

[0006] Based on the above background, existing studies mostly explore low-dimensional hand movement control manifolds in the neural signal space or the motion space, but often ignore the connection between the two, which may lead to a mismatch between the neural manifold and the motion manifold, thereby limiting the decoding effect of high-dimensional hand movements. Summary of the invention

[0007] The present invention provides a dexterous hand movement decoding method based on collaborative basis manifolds. By adopting a joint manifold learning method that takes both neural signals and motion signals into consideration, the guidance and constraints of EEG data are added to the learning of motion collaborative basis, so that the learned collaborative basis manifold can better reflect the actual control mechanism of the brain, thereby achieving more accurate and precise hand movement decoding.

[0008] A dexterous hand motion decoding method based on collaborative basis manifolds, comprising: (1) Obtain EEG signal features and corresponding hand movement data, align the time series, and perform preprocessing to obtain a training data set; (2) Construct a dexterous hand motion decoding model based on the collaborative basis manifold, including a periodic network module, a control network module, and a collaborative basis recruitment module; Among them, the periodic network module is used to map the EEG signal features into the phase manifold to obtain EEG features with periodic characteristics; the control network module extracts the decoupled spatiotemporal motion control parameters from the EEG features obtained by the periodic network module; the synergistic basis recruitment module recruits the corresponding time-varying synergistic basis from the maintained synergistic basis dictionary according to the spatiotemporal motion control parameters obtained by the control network module, and combines and decodes to obtain the corresponding hand movement; (3) Construct an overall loss and use the training dataset to train the dexterous hand movement decoding model; (4) Extract the features of the EEG signal to be predicted, input it into the trained dexterous hand movement decoding model, and decode it to obtain the corresponding hand movement.

[0009] In step (1), the spike frequency band energy in the invasive EEG data is extracted as the EEG signal feature.

[0010] In step (2), the periodic network module is developed based on the autoencoder framework, including: time domain convolutional layer and deconvolution layer, Fourier transform-based periodic encoding layer and decoding layer, and phase manifold mapping layer for periodic motion control feature extraction; the training of the periodic network module includes a reconstruction loss of the neural signal autoencoder.

[0011] In step (2), the control network module includes two independent feedforward network layers, which map the EEG features obtained from the periodic network module into spatiotemporal motion control parameters including amplitude scaling factors and time offsets.

[0012] In step (2), the control network module also includes a decoupling module for decoupling the spatiotemporal motion control parameters during the training phase. The decoupling module obtains motion parameters that independently control the motion time and space by exchanging parameters between different samples in a batch. The training of the control network module includes a decoupling loss.

[0013] In step (2), the synergistic basis recruitment module maintains a synergistic basis dictionary, and selects the corresponding time-varying synergistic basis from the synergistic basis dictionary based on the amplitude scaling factor and time offset obtained by the control network module; the amplitude is scaled using the corresponding amplitude scaling factor, and the time step of the synergistic basis is selected using the time offset, and then the decoding result of the hand movement is obtained by linearly combining all synergistic bases; the training of the synergistic basis recruitment module includes a decoding loss of motion data.

[0014] In step (2), the collaborative basis recruitment module maintains a collaborative basis dictionary. During the training phase, the collaborative basis dictionary is used as a learnable parameter tensor and is used together with other parts of the network to estimate parameters through the back-propagation method.

[0015] The parameter learning process of the collaborative basis dictionary is constrained by the dictionary's regularization terms, including the smoothing regularization term in the time dimension and the orthogonalization regularization term between dictionary elements; the training of the collaborative basis recruitment module includes a regularization loss for dictionary learning.

[0016] In step (3), the overall loss constructed includes the reconstruction loss of the neural signal autoencoder from the periodic network module, the decoupling loss from the control network module, the decoding loss from the collaborative basis recruitment module, and the regularization loss of the dictionary learning.

[0017] In step (3), during the training of the dexterous hand movement decoding model, parameter estimation is performed based on the back propagation method.

[0018] Compared with the prior art, the present invention has the following beneficial effects: Different from the previous methods of constructing neural manifolds or motion manifolds separately, the present invention adopts a joint manifold learning method to take both neural signals and motion signals into consideration. At the neural signal level, a periodic manifold representation is adopted; at the motion signal level, a synergistic basis is introduced to encode multi-joint hand movements. At the same time, the neural representation and the motion representation are associated through motion control parameters, so that the learning process of the hand motion manifold is supervised by the neural signal, thereby realizing the joint manifold learning of the two modal signals. This method makes the hand motion process more consistent with the natural control mode of the brain. Experimental results show that the method of the present invention can achieve more accurate and sophisticated hand motion decoding. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The framework flow chart of a dexterous hand motion decoding method based on collaborative basis manifolds in an embodiment of the present invention.

[0020] Figure 2 FIG. 4 is a diagram of the experimental paradigm of the data set in the implementation of the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0022] All clinical and experimental procedures in the embodiments of the present invention were approved by the Medical Ethics Committee of the Second Affiliated Hospital of Zhejiang University (Ethics Review No. 2019-158, approved on May 22, 2019). The subject was a 74-year-old male who suffered a C4-level cervical spine injury caused by a car accident, quadriplegia, and could only move the part above his neck. He had normal language communication ability and task comprehension ability. Two 96-channel Utah intracortical microelectrode arrays (Blackrock Microsystems, Salt Lake City, Utah, USA) were implanted in the left primary motor cortex of the subject to record neural signals. The subject performed brain-computer interface training tasks every weekday and rested on weekends.

[0023] like Figure 1 As shown, a dexterous hand motion decoding method based on collaborative basis manifolds includes the following steps: Step 1, data preprocessing: obtain the original invasive EEG data recorded by the hardware and the corresponding hand movement data; extract the spike potential frequency band energy in the invasive EEG data as the EEG signal feature for dexterous hand movement decoding; align the EEG signal features with the corresponding hand movement data through a sliding time window to form a time series sample set; divide the data into training set, validation set and test set in a reasonable proportion, standardize the data, and obtain the preprocessed EEG data and the corresponding movement data.

[0024] The experimental paradigm for data collection is as follows Figure 2 As shown in Figure 1. The dataset includes EEG signals and corresponding motion signals collected when a subject watched a virtual hand movement in an instructional video and imagined performing the same hand movement. Each experimental trial of the paradigm consists of two stages: "preparation" and "execution". In the "preparation" stage, the type of action to be performed will be prompted by pictures and voice; then in the "execution" stage, the virtual hand will perform the action 5 times in a row, and the subject will observe and imagine that he is also performing this action. The paradigm includes a total of 11 different hand movements, namely: bending the thumb, bending the index finger, bending the middle finger, bending the ring finger, bending the little finger, bending the back four fingers, bending the back three fingers, five-finger grip, three-finger grip, two-finger grip, and thumb adduction. The execution time of each action is 2.4 seconds; each block will contain all 11 actions in a random order, and each session contains 4 blocks.

[0025] During the execution of the action, the angles of 15 key joints of the hand were recorded as motion data, with a collection frequency of 250 Hz. After smoothing, symmetry processing and upsampling to 1000 Hz, the time series motion data of 11 different actions were obtained. , where M represents the feature dimension of the motion signal, specifically 15, Indicates the time required to complete a complete action, specifically 2400.

[0026] At the same time, two 96-channel Utah intracortical microelectrode arrays implanted in the subjects' cortex were used to synchronously collect neural signals. The Neuroport system (Blackrock Microsystems) was used to amplify, digitize and sample the neural signals at 30 kHz. In order to denoise the signal, high-pass (0.3 Hz) and low-pass (7.5 kHz) Butterworth filters were used to filter the data. The acquisition was carried out on three different experimental days: 1035 days, 1081 days, and 1105 days after implantation.

[0027] The original invasive EEG data obtained by 30k Hz sampling was downsampled to 1000 Hz after 250 Hz high-frequency filtering, 3000-1000 Hz band filtering, and half-wave rectification to obtain the spiking band power (SBP) as the decoded EEG signal feature X. ,in Indicates the characteristic dimension of the EEG signal, specifically 196, Indicates the time required to complete a complete action, specifically 2400.

[0028] The EEG data and motion data are preprocessed using a sliding window. The window length and step size can be set according to the characteristics of the data. In this embodiment, the window length is set to 400ms and the step size is set to 200ms. The data is averaged within the window to obtain data with 11 steps. For the motion signal at each time step, the EEG signals of the previous N time steps are used for prediction. According to the principle of k-fold cross validation, the data is used as the training set, the data of k-2 folds is used as the validation set, and the data of 1 fold is used as the test set, and repeated k times. In each fold of data, the motor data and neural data are standardized according to the mean and variance obtained from the training set. Use The average performance of all validation sets is used as the standard to select the best model. The average performance of the folded test set is taken as the final performance indicator.

[0029] In this embodiment, k=4, and the data is divided into a training set, a validation set, and a test set in a 2:1:1 manner according to a 4-fold cross-validation method.

[0030] Step 2, establish a dexterous hand motion decoding model based on collaborative basis manifold, which mainly includes a periodic network module, a control network module and a collaborative basis recruitment module.

[0031] 2-1. Periodic network module: maps EEG features to phase manifolds to obtain EEG features with periodic characteristics; Specifically, the periodic network module is developed based on the autoencoder framework, which includes time domain convolutional and deconvolutional layers, Fourier transform-based periodic encoding and decoding layers, and phase manifold mapping layers for periodic motion control feature extraction. The specific calculation method is as follows: First, the latent representation is extracted from the neural signal based on the time domain convolution layer: ; in, Indicates that it contains channels, The neural signal of time steps, The dimension is The implicit representation of . represents a one-dimensional temporal convolution layer, whose convolution kernel scale can be set according to the characteristics of the data. is 196, the number of time steps is 5, the dimension of the implicit representation of the number of channels The optimal size is selected from 320 to 512 with an interval length of 32. The optimal size of the time dimension of the one-dimensional temporal convolution layer is selected from {3, 5}.

[0032] Next, the latent representation with temporal features is modeled by a periodic sine function, whose parameters include amplitude ( ),frequency( ), offset( ) and phase shift ( ). In order to estimate these parameters, the latent representation Perform a fast Fourier transform ( ), and obtain the Fourier coefficients ,in Then, according to the coefficient size of each frequency component, each dimension is calculated by weighted average The magnitude of the latent representation ,frequency and offset : ; in Indicates frequency components The power, is the frequency vector. Phase offset Then we learn through two fully connected layers (FC): ; In order to enhance the periodic representation characteristics of latent representation for motion control, a periodic control manifold is defined , mapping the corresponding representation onto the manifold : ; In order to avoid overfitting of the network by relying only on motion data, we add a decoder to the network to reconstruct the original neural signal. Specifically, the latent representation is reconstructed according to the learned parameters: ; in, represents the time window. Then one-dimensional deconvolution is used to convert Mapping back to the original neural signal space: ; In this embodiment, the time dimension size of the one-dimensional de-time convolution layer is consistent with the one-dimensional de-time convolution layer.

[0033] The network learns by minimizing the reconstruction loss function between the original neural signal and the reconstructed neural signal: ; In this embodiment, the mean square error (MSE) is used to measure the loss.

[0034] 2-2. Control network module: extract decoupled spatiotemporal motion control parameters from EEG features obtained based on the periodic network module; Specifically, the control network module contains two independent feedforward network layers, which map the EEG features obtained from the periodic network module into motion control parameters including amplitude scaling factors and time offsets. The specific calculation method is as follows: Two independent feed-forward network (FFN) layers are used to predict the amplitude scaling factor ( ) and time offset ( ), where K represents the number of synergistic bases. Each FFN consists of two layers of linear transformations and the activation function ReLU in between. To encourage the sparsity of activation, the activation function ReLU is added after the feedforward network layer of the amplitude scaling factor; and to represent the time offset in a relative time style, we will The value range of is limited to (0, 1], so the activation function Tanh after scaling and offset transformation is added after the corresponding feedforward network. The calculation method is as follows: ; In this embodiment, the size of the intermediate layer dimension of FFN is optimally selected from {196, 256}.

[0035] In order to encourage the control network module to extract decoupled motion time and space control parameters, the control network contains a parameter decoupling module during the training phase. This module strengthens the independence of spatiotemporal control parameters through a "data augmentation" strategy during training. Specifically, the execution of a single continuous action can be decomposed into two parts: action category (what action to do) and action style (how to perform the action). Then the corresponding neural control parameters should also have similar decoupling properties, that is: for an action instance, the scaling factor It should remain relatively stable throughout the entire continuous action; at the same time, for different types of movements but the same style, the time offset The changing rules of should be basically consistent. In order to achieve these two principles, this module randomly exchanges the motion control parameters corresponding to samples with the same action category or stage label in each batch of training, obtains new spatiotemporal parameter pairs, then performs the corresponding collaborative recruitment process and introduces the corresponding decoding loss to encourage independent control of motion parameters. Specifically, the calculation method is as follows: In the training phase, for each batch of samples (the number of samples is ), randomly swap the corresponding and , and the result after exchange is and , formulated as follows: ; in, represents a random permutation operation constrained by the corresponding labels, A label representing the sport category, The sample size of this batch in this example is Set to 32.

[0036] Then, the collaborative recruitment process is performed again for the randomly exchanged sample pairs to obtain the corresponding motion output and . Compare similarity with real motion data, thereby introducing exchange loss : ; in, represents the collaborative basis recruitment function, Indicates the collaborative basis dictionary maintained by it. In this embodiment, the mean square error (MSE) is used to measure the loss.

[0037] 2-3. Collaborative basis recruitment module: Based on the motion control parameters obtained by the control network module, the corresponding time-varying collaborative basis is recruited from the maintained collaborative basis dictionary, and the corresponding hand movement is obtained by combined decoding.

[0038] Specifically, the synergistic basis recruitment module maintains a synergistic basis dictionary, selects the corresponding time-varying synergistic basis from the synergistic basis dictionary based on the amplitude scaling factor and time offset obtained by the control network, scales the amplitude using the corresponding amplitude scaling factor, selects the time step of the synergistic basis using the time offset, and then obtains the decoding result of the hand movement by linearly combining all synergistic bases. The specific calculation method is as follows: The collaborative base recruitment module maintains a collaborative dictionary , each synergistic basis Depending on the magnitude and time offset Perform scaling and time shift operations separately to obtain the synergy basis after the control parameter transformation, and finally add the contributions of all synergy bases to obtain the motion data output The corresponding representation is the collaborative base recruitment function It can be described in the following form: ; However, the collaborative is discrete in the time dimension, and the time offset To solve this mismatch and recruit synergistic bases in a nearly continuous manner, a linear weighted interpolation method can be used. That is, find The weights of the two nearest time steps are calculated based on the distance between them. . Therefore, the recruitment of the coordination base time can be written as: . Therefore, the corresponding recruitment function can be written as: ; Then, the real motion data is compared for similarity and regression decoding loss is introduced Supervised learning of network parameters: ; In this embodiment, the mean square error (MSE) is used to measure the loss.

[0039] The synergy recruitment module maintains a spatiotemporal synergy dictionary ,in is the number of synergistic bases, is the motion output dimension, is the number of time steps (reflecting the time resolution). During the training phase, this dictionary is used as a learnable parameter tensor and is used together with other parts of the network to estimate parameters through the back propagation method. Set to 15, motion output dimension is 15, the number of time steps Set to 20.

[0040] The collaborative base recruitment module maintains a spatiotemporal collaborative base dictionary. During the training phase, the dictionary parameter learning process is constrained by the dictionary regularization term, including the smoothing regularization term in the time dimension and the orthogonal regularization term between dictionary elements. Specifically, the calculation method is as follows: In order to ensure the smoothness of the synergy basis in the time dimension, a smooth regularization loss is defined To penalize large differences between adjacent time steps, thus encouraging collaboration to transition smoothly in time: ; At the same time, in order to reduce the redundancy between synergistic bases and encourage their orthogonality in the feature space, we Expanding in the space dimension, we get , and then calculate its Gram matrix . Define the orthogonalization regularization loss To minimize With the identity matrix Distance between: ; ; ; in, represents the Frobenius norm.

[0041] Step 3: Based on the losses obtained from each module, all losses are summed up with reasonable weights to obtain the overall loss of the model for network training.

[0042] Specifically, the overall loss module integrates the reconstruction loss of the neural signal autoencoder from the periodic network module, the decoupling loss from the control network module, the decoding loss from the collaborative basis recruitment module, and the regularization loss of the dictionary learning. The specific calculation method is as follows: ; in, , and is a hyperparameter used to balance the impact of various losses on the model and can be set manually. Sample uniformly from 0.85 to 0.95 and select the best one. The best one is selected based on log-uniform distribution sampling from 0.1 to 500. Uniformly sample between 0.1 and 1 to select the best one.

[0043] Step 4: Model training and performance evaluation: Train the network in the training set data, evaluate the model's fit using the validation set data, select the model with the best regression effect, and finally evaluate the model performance in the test set. Specifically, the main parameters of model learning are set as follows: the batch size is set to 32; the Adam algorithm is used to optimize the parameters, the learning rate is set to uniformly sample from 0.001 to 0.01 to select the best model based on the performance of the validation set, and the weight decay is set to 1e-4; the maximum learning round is set to 5000, and the early stopping method is used to reduce the overfitting problem. Specifically, the R on the validation set is calculated after each round. 2 , retain the model with the best performance on the validation set. If R 2 If there is no improvement of more than 1% of the previous performance within 30 rounds, the training is stopped and the final network parameters are obtained.

[0044] In order to verify the effectiveness of the proposed dexterous hand movement decoding method based on synergistic basis manifold (SynergyNet), it was compared with the best long short-term memory neural network (LSTM) commonly used in invasive brain-computer interface decoding tasks, and the coefficient of determination (Coefficient of Determination, ), Correlation Coefficient (CC) and Root Mean Squared Error (RMSE) were used to evaluate the performance (see Table 1). As can be seen from Table 1, the performance of SynergyNet is significantly improved compared with LSTM: The improvement is 8.72%, CC is improved by 3.72%, and RMSE is reduced by 6.80%, indicating that our method has a significant advantage in accurately decoding hand movements.

[0045] Table 1

[0046] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A dexterous hand motion decoding method based on collaborative basis manifolds, characterized in that: include: (1) Obtain EEG signal features and corresponding hand movement data, align the time series, and perform preprocessing to obtain a training data set; (2) Construct a dexterous hand motion decoding model based on the collaborative basis manifold, including a periodic network module, a control network module, and a collaborative basis recruitment module; Among them, the periodic network module is used to map the EEG signal features into the phase manifold to obtain EEG features with periodic characteristics; the control network module extracts the decoupled spatiotemporal motion control parameters from the EEG features obtained by the periodic network module; the synergistic basis recruitment module recruits the corresponding time-varying synergistic basis from the maintained synergistic basis dictionary according to the spatiotemporal motion control parameters obtained by the control network module, and combines and decodes to obtain the corresponding hand movement; (3) Construct an overall loss and use the training dataset to train the dexterous hand movement decoding model; (4) Extract the features of the EEG signal to be predicted, input it into the trained dexterous hand movement decoding model, and decode it to obtain the corresponding hand movement.

2. The method for decoding dexterous hand motion based on collaborative basis manifold according to claim 1, characterized in that: In step (1), the spike frequency band energy in the invasive EEG data is extracted as the EEG signal feature.

3. The method for decoding dexterous hand movements based on collaborative basis manifolds according to claim 1, characterized in that: In step (2), the periodic network module is developed based on the autoencoder framework, including: time domain convolutional layer and deconvolution layer, Fourier transform-based periodic encoding layer and decoding layer, and phase manifold mapping layer for periodic motion control feature extraction; the training of the periodic network module includes a reconstruction loss of the neural signal autoencoder.

4. The method for decoding dexterous hand movements based on collaborative basis manifolds according to claim 1, characterized in that: In step (2), the control network module includes two independent feedforward network layers, which map the EEG features obtained from the periodic network module into spatiotemporal motion control parameters including amplitude scaling factors and time offsets.

5. The method for decoding dexterous hand movements based on collaborative basis manifolds according to claim 4, characterized in that: In step (2), the control network module also includes a decoupling module for decoupling the spatiotemporal motion control parameters during the training phase. The decoupling module obtains motion parameters that independently control the motion time and space by exchanging parameters between different samples in a batch. The training of the control network module includes a decoupling loss.

6. The method for decoding dexterous hand motion based on cooperative basis manifold according to claim 4, characterized in that: In step (2), the synergistic basis recruitment module maintains a synergistic basis dictionary, and selects the corresponding time-varying synergistic basis from the synergistic basis dictionary based on the amplitude scaling factor and time offset obtained by the control network module; the amplitude is scaled using the corresponding amplitude scaling factor, and the time step of the synergistic basis is selected using the time offset, and then the decoding result of the hand movement is obtained by linearly combining all synergistic bases; the training of the synergistic basis recruitment module includes a decoding loss of motion data.

7. The method for decoding dexterous hand motion based on collaborative basis manifold according to claim 6, characterized in that: In step (2), the collaborative basis recruitment module maintains a collaborative basis dictionary. During the training phase, the collaborative basis dictionary is used as a learnable parameter tensor and is used together with other parts of the network to estimate parameters through the back-propagation method.

8. The method for decoding dexterous hand motion based on cooperative basis manifold according to claim 7, characterized in that: The parameter learning process of the collaborative basis dictionary is constrained by the dictionary's regularization terms, including the smoothing regularization term in the time dimension and the orthogonalization regularization term between dictionary elements; the training of the collaborative basis recruitment module includes a regularization loss for dictionary learning.

9. The method for decoding dexterous hand movements based on collaborative basis manifolds according to claim 1, characterized in that: In step (3), the overall loss constructed includes the reconstruction loss of the neural signal autoencoder from the periodic network module, the decoupling loss from the control network module, the decoding loss from the collaborative basis recruitment module, and the regularization loss of the dictionary learning.

10. The method for decoding dexterous hand movements based on collaborative basis manifolds according to claim 1, characterized in that: In step (3), during the training of the dexterous hand movement decoding model, parameter estimation is performed based on the back propagation method.

Images