Central nervous system-based exoskeleton control method and system

Abstract

The application belongs to the technical field of robot control, and discloses a central nervous system based exoskeleton control method and system, which comprises the following steps: collecting central nervous system signals in a human body, and preprocessing the central nervous system signals to obtain a multi-dimensional neural feature set; decoding and analyzing the multi-dimensional neural feature set by using a preset dynamic neural intention decoding model to obtain a human body movement intention; generating joint movement parameters and driving force control instructions of an exoskeleton robot by using a collaborative model; controlling the exoskeleton robot to perform joint movement, and constructing a closed loop feedback to adjust the driving force control instructions in real time. The application effectively avoids information loss or distortion caused by damage to the central nervous system transmission path, improves the adaptability to patients with central nervous system damage, improves the decoding accuracy, reduces the movement synchronization error, improves the control fluency and human-machine coordination of the exoskeleton robot, and ensures the control safety of the exoskeleton robot.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot control, and in particular relates to a central nervous system based exoskeleton control method and system. BACKGROUND

[0002] As an important device for assisting human movement, exoskeleton robots are in urgent need in the field of medical rehabilitation. Traditional exoskeleton control relies on surface electromyography signals and mechanical sensor signals, which are indirect signals of peripheral nerves or movement ends. There is a transmission delay and information loss between these signals and the original movement instructions issued by the central nervous system of the human body, resulting in poor coordination between exoskeleton control and human movement intention. Especially for patients with central nervous system damage, the peripheral signals are weak or distorted, and the control effect is even less ideal.

[0003] Existing control technologies based on neural signals are mostly limited to single brain area signals, without integrating signals of central nervous pathways such as the spinal cord, making it difficult to completely analyze human movement instructions. The neural intention decoding model is static and fixed, and cannot adapt to the dynamic changes of central nervous signals. Moreover, there is a lack of collaborative models considering the central nervous system and the dynamics of exoskeletons, and the control instructions are easy to be out of sync with the human movement rhythm.

[0004] Therefore, how to provide a central nervous system based exoskeleton control method and system is a problem to be solved at present. SUMMARY

[0005] The embodiment of the present application provides a central nervous system based exoskeleton control method and system to solve the problems in the prior art.

[0006] To provide a basic understanding of some aspects of the disclosed embodiments, the following brief summary is given. The summary is not an extensive overview of the application, nor is it intended to identify key / critical elements of the application or to delineate the scope of the embodiments. Its sole purpose is to present some concepts in a simplified form as a prelude to the more detailed description that is to follow.

[0007] According to a first aspect of the embodiment of the present application, a central nervous system based exoskeleton control method is provided.

[0008] In one embodiment, the central nervous system based exoskeleton control method comprises:

[0009] Collecting central nervous system signals in the human body, and preprocessing the central nervous system signals, extracting multi-dimensional neural features according to the preprocessed central nervous system signals, and obtaining a multi-dimensional neural feature set;

[0010] Decoding and analyzing the multi-dimensional neural feature set by using a preset dynamic neural intention decoding model to obtain a human movement intention;

[0011] Based on the human motion intention, a central nervous signal-exoskeleton robot cooperation model is used to generate joint motion parameters and driving force control instructions of the exoskeleton robot;

[0012] Based on the joint motion parameters and driving force control instructions of the exoskeleton robot, the exoskeleton robot is controlled to perform joint motion, and the motion state information and human-machine interaction force information of the exoskeleton robot are collected to construct a closed-loop feedback to adjust the driving force control instructions in real time.

[0013] In one embodiment, the central nervous signal is preprocessed, and multi-dimensional neural features are extracted from the preprocessed central nervous signal to obtain a multi-dimensional neural feature set, which includes:

[0014] The central nervous signal is subjected to noise removal by using an adaptive Kalman filtering algorithm, and physiological artifacts are removed from the central nervous signal after noise removal to obtain a removed central nervous signal;

[0015] The removed central nervous signal is subjected to feature enhancement by combining a wavelet packet decomposition algorithm and a singular value decomposition algorithm, and a reconstruction matrix is constructed;

[0016] Multi-dimensional feature extraction is performed on the reconstruction matrix to obtain a multi-dimensional neural feature set.

[0017] In one embodiment, the removed central nervous signal is subjected to feature enhancement by combining a wavelet packet decomposition algorithm and a singular value decomposition algorithm, and a reconstruction matrix is constructed, which includes:

[0018] Based on a preconfigured decomposition layer number and a target frequency band, sub-band screening is performed on the removed central nervous signal, and threshold denoising processing is performed on the screened sub-band;

[0019] The sub-band after threshold denoising processing is subjected to signal reconstruction by using a wavelet packet decomposition method to obtain a reconstructed signal, and the obtained reconstructed signal is subjected to segmentation processing according to a time axis to obtain an original feature matrix;

[0020] The original feature matrix is subjected to matrix centering processing to obtain a centered matrix;

[0021] The elements in the centered matrix are decomposed and extracted by using a singular value decomposition algorithm to obtain a reconstruction matrix.

[0022] In one embodiment, the multi-dimensional neural feature set is decoded and analyzed by using a preset dynamic neural intention decoding model to obtain a human motion intention, which includes:

[0023] The multi-dimensional neural feature set is input into the preset dynamic neural intention decoding model, and a plurality of feature subsets are obtained by feature division on the multi-dimensional neural feature set;

[0024] dynamically assigning features in each feature subset by using an attention mechanism to obtain attention weights;

[0025] compressing each feature subset by using a multi-layer perception algorithm to obtain a compressed feature vector, and splicing each compressed feature vector to obtain a fusion feature vector by combining the attention weights;

[0026] According to the fusion feature vector, a human motion intention probability distribution is calculated by using a motion intention probability distribution formula, and the human motion intention is determined.

[0027] In one embodiment, the dynamically assigning features in each feature subset by using an attention mechanism to obtain attention weights comprises:

[0028] pooling each feature subset to obtain a channel descriptor, and generating a channel weight by using the channel descriptor to obtain a channel attention weight value;

[0029] extracting a maximum weight value on a feature map in the feature subset by using an attention heat map to obtain a spatial attention weight value;

[0030] calculating the channel attention weight value and the spatial attention weight value to obtain a final attention weight.

[0031] In one embodiment, the pooling each feature subset comprises:

[0032] obtaining time series data of each feature channel in each feature subset, and performing verification and outlier processing on the data to obtain pooled data;

[0033] performing global average pooling and maximum pooling operations on the pooled data based on a pre-configured pooling window and step to obtain a global average value and a global maximum value;

[0034] splicing the global average value and the global maximum value along the feature channel direction to obtain a channel descriptor.

[0035] In one embodiment, the human motion intention formula is:

[0036] ;

[0037] In the formula, represents a multi-dimensional neural feature set, represents a long short-term memory network output, represents an attention mechanism layer, represents an output layer parameter, represents a motion intention category.

[0038] In one embodiment, the generating, based on the human motion intention, the joint motion parameter and the driving force control instruction of the exoskeleton robot by using the central nervous signal-exoskeleton robot coordination model comprises:

[0039] acquiring the body state parameter of the target user, and combining the human motion intention, generating the joint motion parameter of the exoskeleton robot of the target user by using the central nervous signal-exoskeleton robot coordination model;

[0040] calculating the central nervous conduction delay parameter of the target user by using a time compensation algorithm, and generating the driving force control instruction of the exoskeleton robot.

[0041] In one embodiment, the formula of the central nervous signal-exoskeleton robot coordination model is as follows:

[0042] ;

[0043] In the formula, denotes the exoskeleton joint driving moment, denotes the stiffness coefficient, denotes the target joint angle, denotes the actual joint angle, denotes the damping coefficient, denotes the target joint angular velocity, denotes the actual joint angular velocity, denotes the correction force function based on the central nervous signal.

[0044] According to a second aspect of the embodiment of the present application, a central nervous-based exoskeleton control system is provided.

[0045] In one embodiment, the central nervous-based exoskeleton control system comprises:

[0046] a preprocessing module, which collects the central nervous signal in the human body, and pre-processes the central nervous signal, extracts multi-dimensional neural features according to the pre-processed central nervous signal, and obtains a multi-dimensional neural feature set;

[0047] a decoding module, which decodes and analyzes the multi-dimensional neural feature set by using a preset dynamic neural intention decoding model, and obtains the human motion intention;

[0048] a data generation module, which generates, based on the human motion intention, the joint motion parameter and the driving force control instruction of the exoskeleton robot by using the central nervous signal-exoskeleton robot coordination model;

[0049] The adjusting module controls the exoskeleton robot to perform joint movement based on joint movement parameters and driving force control instructions of the exoskeleton robot, collects movement state information and human-machine interaction force information of the exoskeleton robot, constructs a closed-loop feedback, and adjusts the driving force control instructions in real time.

[0050] According to a third aspect of the embodiments of the present application, a computer device is provided.

[0051] In some embodiments, the computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0052] According to a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided.

[0053] In one embodiment, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0054] The technical solutions provided by the embodiments of the present application can have the following beneficial effects.

[0055] (1) The central nervous system signals are comprehensively collected, which are closer to original movement instructions than peripheral signals, effectively avoiding information loss or distortion caused by damage to the central nervous system transmission path, and improving the adaptability to patients with central nervous system damage.

[0056] (2) The dynamic neural intention decoding model adapts to the dynamic changes of the central nervous system signals, and compared with the traditional static model, the decoding accuracy is improved by 15%-20%.

[0057] (3) The central nervous system-exoskeleton collaborative model considers the dynamic characteristics of both sides, and the model not only decodes the user's intention, but also considers the dynamic characteristics of the exoskeleton itself in real time, so that the movement synchronization error is reduced to within 20ms, the user's feeling of drag and resistance is reduced, and the control fluency and human-machine coordination of the exoskeleton robot are improved.

[0058] (4) The closed-loop feedback combines PID and reinforcement learning to quickly compensate for instantaneous physical disturbances, ensure that each movement of the exoskeleton is accurately and stably realized, and balance the adjustment speed and stability to ensure the control safety of the exoskeleton robot.

[0059] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0060] The accompanying drawings, which are incorporated herein and constitute part of this specification, illustrate embodiments consistent with the application and, together with the description, further serve to explain the principles of the application.

[0061] Figure 1 is a flow chart of a central nervous system based exoskeleton control method according to an exemplary embodiment;

[0062] Figure 2 is a principle block diagram of a central nervous system based exoskeleton control system according to an exemplary embodiment;

[0063] Figure 3 is a structural schematic diagram of a computer device according to an exemplary embodiment;

[0064] Figure 4 is a whole flow schematic diagram of a central nervous system based exoskeleton control method in application according to an exemplary embodiment;

[0065] Figure 5 is a structural schematic diagram of a dynamic neural intention decoding model in a central nervous system based exoskeleton control method according to an exemplary embodiment. DETAILED DESCRIPTION

[0066] The following description and drawings are illustrative of specific embodiments of the present document and are not intended to be limiting of the present document. Particular embodiments of the present document will be described below with reference to the accompanying drawings. In the drawings, the size of layers and regions are exaggerated for clarity. Like numbers refer to like elements throughout. The following detailed description is not limiting in any way. The present document can be practiced in other ways than those specifically discussed in the following description and drawings. The present document is not limited to the embodiments described and shown in the drawings, which are provided for merely for explanatory purposes. The drawings are included to provide a practical understanding of certain embodiments of the present document and are not provided to limit the present document. It should be understood that numerous variations and modifications can be made to the embodiments described and shown in the drawings without departing from the present document. The present document can be practiced in other ways than those specifically discussed in the following description and drawings. The drawings are not necessarily to scale; instead, features can be exaggerated to illustrate certain features. The present document is not limited to the specific embodiments described and shown in the drawings, but can include other embodiments that are now known or later developed, including embodiments that do not involve the limitations discussed in the background section of this document. The drawings are included to provide a practical understanding of certain embodiments of the present document and are not provided to limit the present document. It should be understood that numerous variations and modifications can be made to the embodiments described and shown in the drawings without departing from the present document. The present document can be practiced in other ways than those specifically discussed in the following description and drawings. The drawings are not necessarily to scale; instead, features can be exaggerated to illustrate certain features. The present document is not limited to the specific embodiments described and shown in the drawings, but can include other embodiments that are now known or later developed, including embodiments that do not involve the limitations discussed in the background section of this document.

[0067] The terms "longitudinal", "lateral", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like as used herein to indicate orientation or positional relationships based on the orientations or positional relationships shown in the drawings, are for purposes of this description only, and are not intended to indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and is therefore to be interpreted that the application is not limited to the specific orientations or relative positions described and shown in the drawings.

[0068] In this document, the term "a plurality" includes two or more.

[0069] In this document, the character " / " means that the objects before and after it are in an "or" relationship. For example, A / B means A or B.

[0070] In this document, the term "and / or" is a description of an associated relationship between objects, which means that there can be three relationships. For example, A and / or B means: A or B, or, A and B, the three relationships.

[0071] It should be understood that although each step in the flowchart is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified in this document, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the figure can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with other steps or sub-steps or stages of other steps.

[0072] Each module in the device or system of the present application can be realized wholly or partially by software, hardware and their combination. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to call and execute the operations corresponding to the above modules by the processor.

[0073] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0074] Figure 1 And Figure 4An embodiment of the central nervous system based exoskeleton control method of the present application is shown.

[0075] In this alternative embodiment, the central nervous system based exoskeleton control method and system comprises:

[0076] Step S1, collect the central nervous system signals in the human body, and pre-process the central nervous system signals, extract multi-dimensional neural features according to the pre-processed central nervous system signals, and obtain a multi-dimensional neural feature set;

[0077] In this alternative embodiment, the central nervous system signals are removed from noise using an adaptive Kalman filtering algorithm, and the central nervous system signals removed from noise are subjected to physiological artifact rejection, to obtain the central nervous system signals after rejection. The central nervous system signals after rejection are subjected to feature enhancement using a wavelet packet decomposition algorithm and a singular value decomposition algorithm, and a reconstruction matrix is constructed. Multi-dimensional feature extraction is performed on the reconstruction matrix, to obtain a multi-dimensional neural feature set. The central nervous system signals after rejection are subjected to sub-band screening based on a pre-configured decomposition layer number and a target frequency band, and the screened sub-bands are subjected to threshold denoising processing. The sub-bands after threshold denoising processing are subjected to signal reconstruction using a wavelet packet decomposition method, to obtain a reconstructed signal. The reconstructed signal obtained is subjected to segmentation processing according to a time axis, to obtain an original feature matrix. The original feature matrix is subjected to matrix centering processing, to obtain a centered matrix. The elements in the centered matrix are decomposed and extracted using a singular value decomposition algorithm, to obtain a reconstruction matrix.

[0078] It should be noted that the central nervous system signals include brain cortex electrical signals and spinal cord neural electrical signals, etc. The brain cortex electrical signals are collected by a non-invasive sensor, the electrode distribution covers the M1 area and the S1 area of the motor cortex, and the sampling frequency is 5000 Hz. The spinal cord neural signals are collected by an implanted microelectrode array, the implantation position is determined according to the target controlled limb, such as implanting the lumbar spinal cord when controlling the lower limbs, and the sampling frequency is 10000 Hz, to ensure capturing the subtle features of the neural action potential.

[0079] Specifically, for the power frequency interference (50 Hz) and electromyographic artifacts in the central nervous system signals, the filtering parameters are dynamically adjusted, so that the signal-to-noise ratio is improved to more than 40 dB, the adaptive Kalman filtering algorithm is used for noise removal, and the formula is:

[0080] ;

[0081] In the formula, represents the neural signal filtered at the moment, represents the prediction value at the moment, represents the Kalman gain, represents time measurement value, representing a measurement matrix;

[0082] The physiological artifact removal selects an algorithm according to artifact characteristics, such as an electrooculogram artifact (EOG) using an independent component analysis (ICA) algorithm.

[0083] The wavelet packet decomposition algorithm includes the following steps: selecting a db6 wavelet base function to capture central nervous system signals, setting the decomposition layer number to 5 layers, and dividing the corresponding frequency bands, as follows:

[0084] Layer 1: 0-250Hz;

[0085] Layer 2: 0-125Hz / 125-250Hz;

[0086] Layer 3: 0-62.5Hz / 62.5-125Hz / 125-187.5Hz / 187.5-250Hz;

[0087] Layer 4: 0-31.25Hz to 187.5-218.75Hz, etc. 8 sub-bands;

[0088] Layer 5: 0-15.625Hz to 234.375-250Hz, etc. 16 sub-bands.

[0089] In addition, the reconstruction threshold value adopts an adaptive soft threshold value, and the calculation formula based on the standard deviation of each sub-band noise is as follows:

[0090] ;

[0091] In the formula, denotes the signal length, denotes the standard deviation of each sub-band noise.

[0092] The target frequency band is set to 8-500Hz, the effective sub-bands are selected according to the above hierarchical results, and the effective sub-bands are as follows:

[0093] 8-15.625Hz (sub-band 3 of layer 5), 15.625-31.25Hz (sub-band 4 of layer 5), 31.25-62.5Hz (sub-band 2 of layer 4), 62.5-125Hz (sub-band 2 of layer 3), 125-250Hz (sub-band 2 of layer 2), 250-500Hz (obtained by signal mirroring expansion of layer 1), and the sub-band with severe noise is removed. The signal-to-noise ratio (SNR) is , and 6 effective sub-bands are reserved.

[0094] The effective sub-bands are respectively subjected to the denoising processing of the signal components of the 6 frequency bands, and the reconstructed signal is obtained:

[0095] ;

[0096] and the ability value of each component is calculated as the basis for subsequent feature screening, and the formula is:

[0097] ;

[0098] In the formula, represents the ability value, represents the th reconstructed signal;

[0099] Wherein, the steps of singular value decomposition algorithm are:

[0100] 1. Segment the reconstructed signal of the remaining frequency band according to the time axis, and each segment has a length of 1024 points (corresponding to 2.048s, sampling rate 500Hz), to obtain the original feature matrix:

[0101] ;

[0102] In the formula, represents the number of samples, such as the number of collected neural signal segments, represents the feature dimension, such as the total dimension of time domain and frequency domain features, represents a real number set, which is used to describe the number domain attribute of the original feature matrix X, that is, each element in the matrix X is a real number.

[0103] The original feature matrix is centrally processed:

[0104] ;

[0105] In the formula, represents the mean vector of each row, which eliminates the interference of direct current component.

[0106] The element definition formula of is calculated as:

[0107] ;

[0108] In the formula, is the matrix element, which represents the th feature value of the th sample, represents the mean value of the th feature dimension.

[0109] The definition formula of the matrix after centralization processing is:

[0110] ;

[0111] In the formula, represents a column vector of all 1s.

[0112] Furthermore, the specific formula for calculating matrix elements is as follows:

[0113] ;

[0114] 2. Perform singular value decomposition on the centered matrix:

[0115] ;

[0116] In the formula, This represents the left singular matrix, characterizing the correlation between frequency bands. Represents the singular values ​​of a diagonal matrix This represents a right singular matrix, characterizing time dependence.

[0117] The contribution rate of singular values ​​is calculated by selecting the cumulative contribution rate. The first three principal components are given by the following formula:

[0118] ;

[0119] In the formula, This represents the contribution rate of the k-th singular value. Indicates the first k A singular value, Indicates the first A singular value.

[0120] For example, for contribution rate data This indicates that the contribution rate corresponding to the first singular value is 85.2%. This indicates that the contribution rate corresponding to the second singular value is 7.3%. Calculating these three contribution rates, we get 85.2% + 7.3% + 4.1% = 96.6%, which satisfies... .

[0121] for For example, it is merely the index variable of the summation term in the formula, representing the 1st to 6th singular values, i.e. The value range is from 1 to 6.

[0122] 3. Retain the components corresponding to the first three singular values ​​to obtain the reconstructed matrix. The formula is:

[0123] ;

[0124] In the formula, Represents the first three columns of the left singular matrix. Represents the first three columns of a diagonal matrix. This represents the first three columns of the right singular matrix.

[0125] The dimension of the reconstruction matrix is 6x1024, which retains the main features of neural firing, such as the time-domain waveform of action potentials, frequency band energy distribution, and eliminates redundant information caused by noise.

[0126] Three features are extracted from the reconstruction matrix: time-domain features, frequency-domain features, and singular value features.

[0127] The total dimension of the three features is calculated to obtain a multi-dimensional neural feature set, including the event-related desynchronization / synchronization (ERD / ERS) features of the cerebral cortex electrical signal, the firing frequency (5-300Hz) and spike firing pattern features of the spinal cord neural signal, forming a 128-dimensional neural feature vector.

[0128] For example, time-domain features: peak value, kurtosis, and zero-crossing rate of each frequency band signal, 6 frequency bands x 3 features = 18 dimensions; frequency-domain features: normalized value of each frequency band energy , 6 dimensions, energy ratio between frequency bands, such as 4 dimensions, singular value features: the first 3 singular values and their ratios 4 dimensions;

[0129] Total dimension value: 18+6+5+4=33 dimensions.

[0130] 4. Verification of enhancement effect:

[0131] (1) Signal-to-noise ratio improvement: the SNR of the neural signal before processing is 3-8dB, and after processing it is improved to 15-22dB.

[0132] (2) Feature discrimination: the cosine distance of the feature vectors of "motor preparation" and "rest" state increases from 0.35 to 0.72, and the intra-class distance decreases from 0.28 to 0.11.

[0133] (3) Subsequent recognition accuracy: after combining the feature enhancement method, the neural intention recognition accuracy is improved by 8.3% (from 82.5% to 90.8%) compared to single wavelet decomposition.

[0134] For example:

[0135] 1. Patient wears 32-channel high-density EEG device (electroencephalograph), electrodes are positioned according to the international 10-20 system, among which C3, C4 and Cz electrodes focus on covering the M1 area of the motor cortex, the sampling frequency is set to 5000 Hz, the signal resolution is 16 bit, the input impedance is ≥100 MΩ, and the weak cortical electrical activity is ensured to be captured. The lumbar spinal cord (L2-L4 segments) is implanted with a 4x4 microelectrode array with an electrode spacing of 500 μm and a material of platinum-iridium alloy, which is fixed outside the dura mater through minimally invasive surgery, the sampling frequency is 10000 Hz, the bandwidth is 0.1-10 kHz, and the spike signal (amplitude 50-500 μV) of a single neuron can be detected. In the system initialization stage, 5 minutes of resting state data are collected: EEG records the α wave (8-13 Hz) baseline power, and the spinal cord signal records the spontaneous firing frequency (5-15 Hz), and the electrode offset (≤2 μV / h) is eliminated through the reference calibration algorithm, and the individual neural signal template library is established.

[0136] 2. The patient generates the "left leg lifting" motor intention under the guidance of the rehabilitation therapist, at this time the EEG device detects that the M1 area (C4 electrode) appears significant ERD characteristics: the power in the 8-13 Hz band decreases by 30% (from 5 μV² to 3.5 μV²) within 1.2 seconds from the baseline value, accompanied by a temporary increase of 15% in the 18-25 Hz β wave power, forming a typical motor preparation mode of "μ wave suppression + β wave enhancement". The spinal cord microelectrode synchronously records the spike firing frequency of the L3 segment motor neuron from 10 Hz in the resting state to 45 Hz, the firing interval is shortened from 80 ms to 22 ms, and burst firing (continuous 3-5 spike intervals ≤10 ms) is presented, which is consistent with the nerve conduction pathway of the quadriceps muscle activation.

[0137] 3、Call adaptive Kalman filter algorithm, set process noise covariance Q = 1e-6, measurement noise covariance R dynamic adjustment, EEG signal R = 5e-5, spinal cord signal R = 1e-5. The electrooculogram (EOG) in the EEG signal is separated by independent component analysis (ICA) and removed, and the signal-to-noise ratio of the processed signal is improved from 28dB to 42dB; for the electromyographic interference (200-500Hz) in the spinal cord signal, 50Hz notch filter and wavelet threshold denoising are combined, so that the spike signal detection rate reaches 98%. Three types of features are extracted from the EEG signal: ERD / ERS power change rate (8-13Hz, 18-25Hz), event-related potential (ERP) latency, time from intention generation to signal peak 280ms; cross-electrode coherence (coherence coefficient between C3-C4 electrodes 0.65); four types of features are extracted from the spinal cord signal: spike firing frequency (45Hz), firing interval standard deviation (8ms), burst event proportion (35%), waveform similarity (matching degree with template spike 92%); after dimension reduction by principal component analysis (PCA), a 128-dimensional feature vector is formed, retaining 95% of the original information.

[0138] Step S2, decoding and analyzing the multi-dimensional neural feature set by using a preset dynamic neural intention decoding model to obtain a human motion intention;

[0139] In this optional embodiment, the input multi-dimensional neural feature set is divided into several feature subsets, each feature subset is compressed by using a multilayer perception algorithm to obtain a compressed feature vector, and each compressed feature vector is spliced to obtain a fusion feature vector; the delay relationship of the fusion feature vector is captured by using a recurrent neural network algorithm, and each feature subset is dynamically weighted by using an attention mechanism algorithm.

[0140] Each feature subset is pooled to obtain a channel descriptor, and the channel descriptor is used to generate a channel weight to obtain a channel attention weight value; the maximum weight value on the feature map is extracted by using an attention heat map to obtain a spatial attention weight value; the channel attention weight value and the spatial attention weight value are calculated to obtain a final attention weight. The data to be pooled is obtained, and the data is verified and processed for abnormal values; the pooling window and the step are uniformly set, the pooled data is calculated to obtain global average pooling and maximum pooling; the global average pooling and the maximum pooling are spliced to obtain a channel descriptor.

[0141] The output human motion intention probability value formula is:

[0142] ;

[0143] In the formula, denotes a multi-dimensional neural feature set, denotes a long short-term memory network output, denotes an attention mechanism layer, denotes an output layer parameter, denotes a motion intention class.

[0144] It should be noted that the recurrent neural network (LSTM) captures the time sequence dependence of the central nervous signal, such as the transmission delay relationship (about 20-50 ms) between the cerebral cortex signal and the spinal cord signal; the attention mechanism layer gives dynamic weights to different neural features, such as the cerebral cortex signal in the motion preparation stage (weight ratio 70%) and the spinal cord signal in the motion execution stage (weight ratio 60%); through the human motion intention probability value formula, 6 basic motion intentions (such as hip / knee / ankle flexion and extension, overall walking, etc.) are identified, and the decoding accuracy is 92.3%, and the response time is ≤150 ms.

[0145] Specifically, as shown in Figure 5 , the multi-dimensional neural feature set is divided into three feature subsets, namely the cerebral cortex electrical signal feature (48 dimensions), the spinal cord neural signal feature (64 dimensions), and the cross-correlation feature (16 dimensions). Each feature subset is mapped by an independent multi-layer perceptron (MLP) for feature mapping and compressed to 32 dimensions, and then spliced into 96 dimensions.

[0146] In the formula, the cross-correlation feature is obtained by calculating the cross-correlation coefficient of the cortex and the spinal cord signal, reflecting the cooperative relationship between the two in time.

[0147] The cerebral cortex feature subset (48 dimensions) is mapped by the MLP mapping formula as follows:

[0148] ;

[0149] In the formula, denotes a 48x32 weight matrix, denotes a 32-dimensional bias vector, denotes a LeakyReLU activation function (slope 0.1).

[0150] The spinal cord feature subset (64 dimensions) is mapped by the MLP mapping formula as follows:

[0151] ;

[0152] In the formula, denotes a 64x32 matrix, denotes a 32-dimensional bias, and the slope of the activation function is set to 0.2 (to enhance the response to spike signals).

[0153] The cross-correlation feature (16 dimensions) is calculated by the formula as follows:

[0154] ;

[0155] wherein, is the cortical signal, is the spinal signal, is the time offset.

[0156] The final fusion feature formula is:

[0157] ;

[0158] wherein, denotes the cerebral cortex, denotes the spinal cord, denotes the cross-correlation.

[0159] 1. The gating mechanism optimization of the recurrent neural network algorithm is that the forget gate, input gate and output gate use adaptive activation functions. For example, when the input feature is the spinal spike firing frequency, the slope parameter of the activation function is increased by 1.5 times, which improves the sensitivity to high-frequency signal changes; when the input feature is the cortical slow wave signal, the slope parameter is reduced to 0.8 times, which enhances the ability to capture low-frequency trends.

[0160] Taking the "left leg lifting" intention as an example, the forgetting coefficient of the forget gate for weak features (spike frequency 10 Hz) 100 ms ago is 0.8 (fast forgetting), and the retention coefficient for current strong features (45 Hz) is 0.9 (key retention). The acceptance coefficient of the input gate for new features increases with the increase of spike frequency, from 0.3 to 0.8.

[0161] The adaptive activation function formula of the forget gate is:

[0162] ;

[0163] wherein, , denotes the slope parameter (for example: spinal feature , cortical feature ), denotes the scaling factor.

[0164] The input gate calculation formula is:

[0165] ;

[0166] The cell state update calculation formula is:

[0167] ;

[0168] The output gate formula is:

[0169] ;

[0170] For example, take the left leg lifting as an example: when (standardized to 1.2), , .

[0171] 2, the calculation process of attention mechanism algorithm is:

[0172] (1) The first stage is channel attention, which realizes the importance weight of each feature channel by squeezing-excitation method, and the global average pooling of each channel feature is obtained. Channel descriptor; Generate channel weight through MLP, for example: the weight of the spinal cord spike frequency channel is 0.9 (the highest), and the weight of the cortex alpha wave power channel is 0.6.

[0173] The steps of global average pooling are:

[0174] Explicit pooling unit: for the 6 frequency band signal components obtained by wavelet packet decomposition and singular value decomposition ( ), each frequency band corresponds to an independent channel, that is, channel 1 corresponds to 8-15.625Hz, channel 2 corresponds to 15.625-31.25Hz, …, and channel 6 corresponds to 250-500Hz; The feature data of each channel is the time series corresponding to the row in the matrix reconstructed by singular value decomposition, with a dimension of 1x1024 (1 row represents 1 channel, and 1024 columns represent time points, with a sampling rate of 500Hz and a time length of 2.048s).

[0175] Data verification and outlier processing:

[0176] The criterion is used to detect outliers in each channel time series. If the feature value of a time point satisfies , , indicates the mean of the channel sequence, , indicates the standard deviation, then replace the outlier with the mean of the channel , if there are missing values in each channel time series, use linear interpolation to fill in the missing values, the interpolation formula is:

[0177] ;

[0178] Set the pooling window and step: use a global single window to cover the entire time series, and the window size is equal to the time length of the channel feature, that is, for a 1024-point time series, the window size is set to 1024, and the step is set to 1024 (consistent with the window size), which ensures that one pooling operation can cover the entire sequence, avoiding overlapping calculation.

[0179] Pooling calculation process: summarize the time series of each channel: And calculate the global average value, the formula is:

[0180] ;

[0181] In the formula, represents the global average pooling result of the th channel, which is a single scalar value.

[0182] For example, the time series of channel 3 (31.25-62.5Hz) is ; Its global average value is:

[0183] ;

[0184] Combine the scalar obtained after outputting each channel to get the pooling feature vector:

[0185] ;

[0186] The dimension is 1x6, which represents the average energy level of each frequency band (channel) in the entire time window, filters the local fluctuations in the time dimension, and retains the overall difference characteristics between channels.

[0187] Pooling and concatenation: The global average pooling result is used as a supplementary feature, combined with the original 33-dimensional feature set (18-dimensional in time domain + 11-dimensional in frequency domain + 4-dimensional in singular value), forming a 39-dimensional extended feature set; The fusion method is feature concatenation, that is, the extended feature vector , where represents the original 33-dimensional feature, represents the 6-dimensional pooling feature.

[0188] Pooling effect verification: The 1024-dimensional time series of each channel is compressed into a 1-dimensional scalar, which retains the core information of the channel while reducing the computational complexity of the subsequent model, the feature dimension increases from 33 to 39, but avoids redundant calculation of time series; The global average pooling has a smoothing effect on random noise, and the feature standard deviation of each channel in the experiment is reduced by 40%-55% compared with before pooling; After adding the pooling feature, the neural intention recognition accuracy is improved from 90.8% to 92.1%, especially for low signal-to-noise ratio (SNR=5-8dB) neural signals, the recognition stability is improved more significantly, and the fluctuation range is reduced from ±3.2% to ±1.8%.

[0189] Specifically, the pooling result is preprocessed to confirm the pooling form, and after average pooling processing, the feature map is from the original dimension , where is the height, is the width, Convert the number of channels to The vector, denoted as . Each element corresponds to the global average value of a single channel; after max pooling, the feature map is also transformed into... The vector, denoted as . Each element corresponds to the global maximum value of a single channel.

[0190] For example: if the original feature map is Then, both average pooling and max pooling yield the following results. The vector. Dimensionality consistency check: Check and Are the number of channels consistent (both are)? If there are discrepancies, such as due to incorrect pooling parameter settings, the original feature map channel count should be used as the standard, and the shorter vector should be zero-padded or the longer vector should be truncated; ensure that the data types of both are consistent (both are float32) to avoid splicing failure due to type mismatch.

[0191] Select splicing dimension: Use channel-based splicing, and from Expand to spliced ​​vector The formula is:

[0192] ;

[0193] In the formula, The concatenated vector represents the first... One element, Indicates the first Each element.

[0194] Perform the concatenation operation: when a matrix is ​​represented, for Column vector, for Column vectors, concatenated to form Column vector, the formula is:

[0195] ;

[0196] For example: when hour, and All Vectors, after concatenation for The vector has the first 64 elements from average pooling and the last 64 elements from max pooling.

[0197] Concatenation dimensionality transformation: transforming the concatenated vectors The dimensions are reshaped to be suitable for subsequent attention calculation, usually one-dimensional (flat to a one-dimensional vector) as the input of the fully connected layer; if the subsequent network requires a specific dimension (such as ), it is adjusted by the reshape operation to ensure that the data form matches the input requirements of the network layer.

[0198] The features after concatenation are subjected to attention mechanism: the concatenated vector is input into the fully connected layer containing the activation function (such as is the weight matrix of , is the compression ratio), and the fused features are extracted after dimension reduction and nonlinear transformation: , and the attention weights are output through the second fully connected layer (the weight matrix of is . where is the weight vector of , used to weight each channel of the original feature map.

[0199] Concatenation effect verification: average pooling captures the global mean of the channel, reflecting the overall intensity, and max pooling captures the peak value of the channel, reflecting the local significant features. After adopting the concatenation strategy, the recognition accuracy of the key channel by the attention mechanism is improved by 5.2%, and the response delay of the exoskeleton control command is reduced by 8ms.

[0200] (2) The second stage is spatial attention, which generates an attention heat map on the feature map. The weight values corresponding to the motor cortex M1 area and the spinal cord L3 segment in the heat map are significantly higher than those in other areas.

[0201] The weight product of the two-stage attention is the final attention weight, which makes the dynamic neural intention decoding model focus on the key features and key brain areas / spinal cord segments.

[0202] Specifically, the channel attention calculation formula is:

[0203] ;

[0204] In the formula, represents global average pooling, represents squeeze-and-excitation operation, represents ReLU function, represents channel weight, represents squeeze layer weight matrix, represents excitation layer weight matrix, represents Sigmoid activation function, which is used to map the result of the squeeze-and-excitation operation to the interval [0, 1] to obtain the channel attention weight , to quantify the importance of different channels.

[0205] For example: spinal spike channel , cortex Wave channel .

[0206] The spatial attention calculation formula is:

[0207] ;

[0208] In the formula, Sigmoid activation function is used to map the feature map processed by the convolution layer To [0, 1] interval, generate spatial attention weight map , highlight the importance of key spatial positions in the feature map.

[0209] After average pooling and maximum pooling of the feature map, the output spatial weight map is obtained by convolution layer .

[0210] For example: M1 region position Segment .

[0211] The final attention weight calculation formula is:

[0212] ;

[0213] For example: total weight of cortical features , spinal cord features .

[0214] Decoding confidence evaluation and error correction mechanism: the decoding output not only contains the movement intention category, but also carries a confidence score, which is calculated by weighting three indicators: the peak value of the model output probability, the weight is 0.4, the matching degree of features and training samples, the weight is 0.3, and the signal-to-noise ratio of neural signals, the weight is 0.3.

[0215] When the confidence score is greater than 0.8, the result is directly output; when 0.6≤score≤0.8, the error correction mechanism is started, the historical decoding results are referred to, if 2 of the last 3 times are the same intention, then confirm the result; when the score is less than 0.6, trigger the signal stream process of reacquisition, the confidence score formula is:

[0216] ;

[0217] In the formula, Indicates the highest intention probability, Indicates the feature matching degree, Indicates the template feature, represents a neural signal feature to be evaluated at present, represents a signal-to-noise ratio score, represents a signal-to-noise ratio of a neural signal, represents a template feature.

[0218] For example: the probability of lifting the left leg is 0.85, the feature matching degree is 0.75, the signal-to-noise ratio is 42dB, and the corresponding score is 0.9.

[0219] Comprehensive confidence score , directly confirm the result.

[0220] For example:

[0221] (1) Feature vector input dynamic decoding model, LSTM layer (3 layers, 64 neurons per layer) captures time correlation through gating mechanism: the first layer extracts instantaneous features within 100ms, the second layer integrates firing patterns within 500ms, and the third layer learns the intention evolution law on a 1.5 second scale, outputting 64-dimensional time sequence features.

[0222] (2) Attention mechanism layer gives dynamic weight to features: in the movement preparation stage (0-500ms), the EEG feature weight accounts for 70%, among which, Wave power feature weight 0.35; in the movement execution stage (500-1500ms), the spinal cord feature weight rises to 60%, and the spike frequency feature weight is 0.4.

[0223] (3) The fully connected layer outputs the probability distribution of 8 categories of intentions: "left leg lifting" probability 0.92, "right leg lifting" probability 0.05, and "still" probability 0.03. After softmax normalization, the intention of "left leg lifting" is determined, and the decoding delay is 180ms from the appearance of ERD features to the output of the intention.

[0224] Step S3, based on the human motion intention, a central nervous signal-exoskeleton robot collaborative model is used to generate joint motion parameters and driving force control instructions of the exoskeleton robot;

[0225] In this optional embodiment, the body state parameters of the target user are acquired, and the joint motion parameters of the exoskeleton robot of the target user are generated through the central nervous signal-exoskeleton robot collaborative model in combination with the human motion intention; the central nervous conduction delay parameters of the target user are calculated by using a time compensation algorithm, and the driving force control instructions of the exoskeleton robot are generated, and the formula of the central nervous signal-exoskeleton robot collaborative model is:

[0226] ;

[0227] In the formula, represents the exoskeleton joint driving torque, represents a stiffness coefficient, represents a target joint angle, represents an actual joint angle, represents a damping coefficient, represents a target joint angular velocity, represents an actual joint angular velocity, represents a correction force function based on the central nervous signal.

[0228] Specifically, the three-dimensional matrix dimensions are: spinal cord nerve spike firing frequency interval, cerebral cortex signal intensity level, and current exoskeleton movement speed.

[0229] For example:

[0230] When the spike firing frequency is 30-50Hz, it is a strong intention. When the cerebral cortex signal intensity level is high, the ERD / ERS change rate is >40%, and the exoskeleton movement speed is <0.3m / s, the corresponding stiffness coefficient K in the parameter matrix is 1.3 times the base value, and the damping coefficient B is 0.8 times the base value;

[0231] When the spike firing frequency is <15Hz, it is a weak intention. When the cortex signal intensity level is low and the movement speed is >0.5m / s, K is 0.6 times the base value, and B is 1.3 times the base value. Through this multi-dimensional parameter adjustment, the synergy model can more accurately match the central nervous instructions in different movement states.

[0232] In the formula, the quantization of the three-dimensional matrix is:

[0233] Spinal cord spike frequency interval: Corresponding quantization value ;

[0234] Cortex signal intensity level: ERD / ERS change rate Corresponding quantization value ;

[0235] Exoskeleton movement speed: Corresponding quantization value ; The stiffness coefficient and damping coefficient adjustment formula is:

[0236] ;

[0237] In the formula, is the base parameter, represents the spinal cord spike frequency interval, represents the cortex signal intensity level, represents the exoskeleton movement speed.

[0238] For example: when Time:

[0239] .

[0240] The dynamic prediction of conduction delay includes the steps of detecting the complexity of the brain cortex signal by sample entropy, when it is increased, it indicates that the neural activity is more active, and the conduction speed can be increased to 110-120m / s; when the complexity of the cortex signal is reduced, , it is reduced to 60-80m / s; the dynamic prediction formula is:

[0241] ;

[0242] In the formula, , the sample entropy of the cortex signal is normalized to [0, 1], in the resting state , in the motion state , in the motion state , in the motion state .

[0243] The conduction time formula is:

[0244] ;

[0245] In the formula, the time unit is ms, , and the conduction distance is m.

[0246] At the same time, a feedback correction mechanism for delay compensation is established, and the time difference between actual motion and neural command is calculated every 50ms, if the difference is more than 10ms, the compensation time of the next time is corrected, and the correction formula is:

[0247] ;

[0248] In the formula, , the compensation time of the n-th time, , the actual synchronization error of the n-th time, , the target error, , and the correction coefficient. For example, if the target error is 5 , the correction coefficient is 0.2, and the actual time difference obtained by calculation is 18ms, which is greater than the target synchronization difference 5ms, then

[0249] , through this dynamic correction, the time synchronization error is stabilized within ±5ms. ms The force-position hybrid strategy is adopted to ensure accurate tracking of joint angles and reasonable distribution of inter-joint forces.

[0250] The force-position hybrid strategy is adopted to ensure accurate tracking of joint angles and reasonable distribution of inter-joint forces. ​

[0251] Among them, for the hip and knee joints, position control is mainly used, force control is auxiliary, the tracking accuracy requirement of position control is ±0.5°, when the joint torque exceeds 80% of the maximum bearing capacity of the joint, the force control mode is automatically switched, and the joint angle is adjusted to reduce the torque;

[0252] For the ankle joint, force control is mainly used, position control is auxiliary, and the contact force between the ankle joint and the ground is ensured to be stable in the preset range, such as 15%-20% of the body weight, when the contact force deviates from the range, the size of the force is corrected by fine-tuning the ankle joint angle.

[0253] In addition to the joint angle coupling formula, a coupling relationship of joint torque is established, and the formula is:

[0254] ;

[0255] In the formula, represents a comprehensive correction term based on the multi-segment neural signals of the spinal cord, which is calculated by weighting the spike signals of L2, L3 and L4 segments, and the weights are 0.3, 0.5 and 0.2 respectively.

[0256] In addition, the extended formula of the coupling relationship of joint torque is:

[0257] ;

[0258] In the formula, represents the spike frequency standardization value (0-1).

[0259] The position control tracking error formula is:

[0260] ;

[0261] In the formula, represents the target joint angle, represents the actual joint angle.

[0262] When , the position controller outputs: ;

[0263] When , switch to force control: ;

[0264] The ankle joint contact force control formula is:

[0265] ;

[0266] In the formula, represents the basic force, which is 17.5% of the body weight, The spike frequency normalized value of the L4 and L2 segments of the spinal cord.

[0267] For example: (1) Call the central nervous system-exoskeleton collaborative model, according to the patient's body weight of 65 kg, look up the table to get the basic stiffness of the hip joint , the damping coefficient . The target angle , the hip joint flexion angle, the target angular velocity , the initial driving torque .

[0268] (2) Time compensation algorithm calculates the nerve conduction delay: the distance from the M1 area to the L3 spinal cord is about 60 cm, the conduction velocity is taken as 80 m / s, the delay time t1=0.6 / 80=7.5 ms; the mechanical response delay of the exoskeleton t2=30 ms, the total compensation time t=37.5 ms, the command is sent 37 ms in advance to ensure that the exoskeleton movement is synchronized with the neural command.

[0269] Step S4, based on the joint motion parameters and driving force control instructions of the exoskeleton robot, control the exoskeleton robot to perform joint motion, and collect the motion state information and human-machine interaction force information of the exoskeleton robot, and construct a closed loop feedback to adjust the driving force control instruction in real time;

[0270] Specifically, since the central nervous system has plasticity, during long-term use, the collaborative model will gradually adapt to the changes in user neural signal characteristics. The system can rotate the user's neural signal pattern once a week for analysis, and if it is found that the average spike firing frequency corresponding to a certain type of motion intention has changed by more than 15% from the initial value, the adjustment coefficient of the collaborative model parameters corresponding to the intention, such as K and B, is fine-tuned, and the fine-tuning amplitude is 5%-10% of the current parameter, so that the collaborative model gradually adapts to the long-term changes in neural signals.

[0271] For example, after a period of training, the average spike firing frequency corresponding to the "left leg lifting" intention decreases from 45 Hz to 38 Hz, and the system fine-tunes the adjustment coefficient of K under this intention from 1.2 (strong intention) to 1.1 to match the change in neural signal intensity.

[0272] Wherein, if the spike frequency average change rate judgment formula of a certain intention is:

[0273] ;

[0274] In the formula, represents the average spike frequency of a certain intention at the current time, represents the average spike frequency of a certain intention at the initial time.

[0275] When the rate of change is > 15%, the cooperative adaptation parameter adjustment formula is:

[0276] ;

[0277] In the formula, represents the adjustment amount of the joint stiffness parameter, represents the adjustment amount of the joint damping parameter, represents the sign function, when the frequency is reduced is negative (reduce stiffness), is positive (increase damping), represents the current joint stiffness parameter, represents the current joint damping parameter.

[0278] In addition, when , the user's motion intention intensity or neural control mode changes, and the joint characteristics need to be dynamically adjusted through the following rules:

[0279] When the frequency is reduced is negative, reduce the current stiffness , make the joint more compliant to adapt to weaker motion intention; is positive, increase the current damping , enhance the motion stability to compensate for the decline in intention intensity.

[0280] When the frequency is increased , is positive, increase the current stiffness, improve the response speed to match stronger motion intention; is negative, reduce the current damping, reduce the motion resistance to support more flexible action.

[0281] The joint angle is collected by the optical encoder on the exoskeleton robot, with a collection accuracy of 0.1°, and the human-exoskeleton robot interaction force is collected by a six-axis force sensor, with a collection accuracy of 0.1N. The collected information is input into the proportional-integral-derivative (PID) and reinforcement learning adjustment algorithm. The state value function update formula of reinforcement learning is:

[0282] ;

[0283] In the formula, represents the current state, represents the learning rate, represents the feedback reward value, represents the discount factor, represents the next state.

[0284] When the joint angle deviation is greater than 5°, it is corrected by PID, and when the interaction force is greater than the threshold value 15N, the interaction force is stabilized in the range of 8N-12N by the function update formula.

[0285] For example:

[0286] (1) The optical encoder of the exoskeleton hip joint (resolution 16384 lines / turn) collects the angle in real time, and the data is refreshed every 10 ms. When moving to 1.2 seconds, the actual angle θ = 32° is detected, and the deviation , the PID module starts to adjust: the proportional term output ;

[0287] The integral term output ( , the integral time is 0.5 s) , the total adjustment torque is 0.2626 N·m, so that the angle converges to 34.5° (deviation 0.5°) within 200 ms.

[0288] (2) The six-axis force sensor (range 0-50N, accuracy 0.1N) detects that the interaction force of the hip joint reaches 14N, triggering reinforcement learning adjustment: the state s is defined as (angle deviation 0.5°, interaction force 14N), and the reward value (reward decreases when deviating from the safety interval 10±2N);

[0289] Update the state value according to the formula , and finally output the torque to 11N, and the interaction force stabilizes at 11.5N.

[0290] (3) The patient trains for 30 minutes every day, and the system triggers incremental learning every 10 minutes: 90% of the historical model parameters are retained, and 10% of the weights are fine-tuned with 500 new samples collected, so that the model adapts to the drift of neural signals, such as the decrease of spike firing frequency by 5-8 Hz with fatigue. After continuous training for 2 weeks, the recognition accuracy of the decoding model for the "left leg lifting" intention is improved from 92% to 95%, the response delay is shortened from 180 ms to 150 ms, and the interaction force adjustment fluctuation range is reduced to ±0.8N, meeting the precise control requirements of rehabilitation training.

[0291] Figure 2 An embodiment of the central nervous system-based exoskeleton control system of the application is shown.

[0292] In this alternative embodiment, the central nervous system-based exoskeleton control system comprises:

[0293] A preprocessing module 1 acquires central nervous system signals in the human body and pre-processes the central nervous system signals, extracts multi-dimensional neural features from the pre-processed central nervous system signals, and obtains a multi-dimensional neural feature set;

[0294] A decoding module 2 decodes and analyzes the multi-dimensional neural feature set using a pre-set dynamic neural intention decoding model to obtain a human motion intention.

[0295] The data generation module 3 generates joint motion parameters and driving force control instructions of the exoskeleton robot based on the human motion intention and by using a central nervous signal-exoskeleton robot cooperation model.

[0296] The adjustment module 4 controls the exoskeleton robot to perform joint motion based on the joint motion parameters and driving force control instructions of the exoskeleton robot, collects motion state information and human-robot interaction force information of the exoskeleton robot, and constructs a closed-loop feedback to adjust the driving force control instructions in real time.

[0297] In an embodiment, a computer device, which can be a server, has an internal structure as shown in Figure 3 The computer device includes a processor, a memory and a network interface connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store static information and dynamic information data. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement the steps in the above method embodiments.

[0298] Those skilled in the art can understand that Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the present application scheme, and does not constitute a limitation on the computer device to which the present application scheme is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0299] In addition, the present application also provides a computer device including a memory and a processor. The memory stores a computer program. The processor implements the steps in the above method embodiments when executing the computer program.

[0300] In addition, the present application also provides a computer readable storage medium having a computer program stored thereon. The computer program is executed by the processor to implement the steps in the above method embodiments.

[0301] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments of each method. In the embodiments of the present application, any reference to memory, storage, database or other medium can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0302] The present application is not limited to the structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present application is only limited by the appended claims.

Claims

1. A central nervous system based exoskeleton control system, characterized by, The system comprises: a preprocessing module, configured to collect central nervous system signals in a human body, and to preprocess the central nervous system signals, extract multi-dimensional neural features according to the preprocessed central nervous system signals, and obtain a multi-dimensional neural feature set; wherein the preprocessing module, when collecting the central nervous system signals in the human body, preprocessing the central nervous system signals, extracting the multi-dimensional neural features according to the preprocessed central nervous system signals, and obtaining the multi-dimensional neural feature set, comprises: using an adaptive Kalman filtering algorithm to remove noise from the central nervous system signals, and removing physiological artifacts from the central nervous system signals after noise removal to obtain central nervous system signals after artifact removal; combining a wavelet packet decomposition algorithm and a singular value decomposition algorithm to enhance the features of the central nervous system signals after artifact removal and construct a reconstruction matrix; extracting multi-dimensional features from the reconstruction matrix to obtain a multi-dimensional neural feature set; a decoding module, configured to decode and analyze the multi-dimensional neural feature set using a preset dynamic neural intention decoding model to obtain a human motion intention; a data generation module, configured to generate joint motion parameters and driving force control instructions of an exoskeleton robot based on the human motion intention and using a central nervous system signal-exoskeleton robot collaboration model; an adjustment module, configured to control the exoskeleton robot to perform joint motion based on the joint motion parameters and driving force control instructions of the exoskeleton robot, collect motion state information and human-machine interaction force information of the exoskeleton robot, and construct a closed-loop feedback to adjust the driving force control instructions in real time.

2. The central nervous system based exoskeleton control system of claim 1, wherein, The preprocessing module, when combining the wavelet packet decomposition algorithm and the singular value decomposition algorithm to enhance the features of the central nervous system signals after artifact removal and construct a reconstruction matrix, comprises: based on a preconfigured decomposition layer number and a target frequency band, performing sub-band screening on the central nervous system signals after artifact removal, and performing threshold denoising processing on the screened sub-bands; using a wavelet packet decomposition method to reconstruct the sub-bands after threshold denoising processing to obtain reconstruction signals, segmenting the obtained reconstruction signals according to a time axis to obtain an original feature matrix; performing matrix centering processing on the original feature matrix to obtain a centered matrix; using a singular value decomposition algorithm to decompose and extract elements in the centered matrix to obtain a reconstruction matrix.

3. The central nervous system based exoskeleton control system of claim 1, wherein, The decoding module, when decoding and analyzing the multi-dimensional neural feature set using the preset dynamic neural intention decoding model to obtain a human motion intention, comprises: inputting the multi-dimensional neural feature set into the preset dynamic neural intention decoding model, dividing the multi-dimensional neural feature set into a plurality of feature subsets, and obtaining a plurality of feature subsets; using an attention mechanism to dynamically assign features to each feature in the feature subsets to obtain attention weights; using a multilayer perception algorithm to compress each feature subset to obtain a compressed feature vector, and combining the attention weights to splice each compressed feature vector to obtain a fusion feature vector; using a motion intention probability distribution formula to calculate a human motion intention probability distribution according to the fusion feature vector, and determining a human motion intention.

4. The central nervous system based exoskeleton control system of claim 3, wherein, The decoding module, when using an attention mechanism to dynamically assign features to the fusion features to obtain attention weights, comprises: The channel description sub is obtained by performing pooling processing on each feature subset, and the channel weight is generated by using the channel description sub, so as to obtain the channel attention weight value; The spatial attention weight value is obtained by extracting the maximum weight value on the feature map in the feature subset by using the attention heat map; The final attention weight is obtained by calculating the channel attention weight value and the spatial attention weight value.

5. The central nervous system based exoskeleton control system of claim 4, wherein, The decoding module comprises the following steps when performing the pooling processing on each feature subset: Time sequence data of each feature channel in each feature subset is obtained, and the data is checked and processed for abnormal values to obtain pooling data; Based on the pre-configured pooling window and step, the global average pooling and the maximum pooling operation are respectively performed on the pooling data to obtain the global average value and the global maximum value; The global average value and the global maximum value are spliced in the feature channel direction to obtain the channel description sub.

6. The central nervous system based exoskeleton control system of claim 5, wherein, The human motion intention formula is: ； wherein, represents a multi-dimensional neural feature set, represents a long short-term memory network output, represents an attention mechanism layer, represents an output layer parameter, represents a motion intention class.

7. The central nervous system based exoskeleton control system of claim 1, wherein, The data generation module comprises the following steps when generating the joint motion parameter and the driving force control instruction of the exoskeleton robot by using the central nervous signal-exoskeleton robot collaborative model based on the human motion intention: The body state parameter of the target user is obtained, and the joint motion parameter of the exoskeleton robot of the target user is generated by combining the human motion intention and the central nervous signal-exoskeleton robot collaborative model; The central nervous conduction delay parameter of the target user is calculated by using the time compensation algorithm, and the driving force control instruction of the exoskeleton robot is generated.

8. The central nervous system based exoskeleton control system of claim 7, wherein, The formula of the central nervous signal-exoskeleton robot collaborative model is: ； In the formula, represents an exoskeleton joint driving moment, represents a stiffness coefficient, represents a target joint angle, represents an actual joint angle, represents a damping coefficient, represents a target joint angular velocity, represents an actual joint angular velocity, represents a correction force function based on the central nervous signal.

Images