An online wrist torque estimation method based on neural features and LSTM

Abstract

The present invention discloses an online wrist torque estimation method based on neural features and LSTM, comprising the following steps: (1) the experimenter keeps the arm still and applies torque to the torque sensor through the wrist; (2) synchronously collects the torque sensor data and high-density surface electromyography signals; (3) uses a blind source separation algorithm to decompose the high-density surface electromyography signals to obtain a motor unit spike train MUST; (4) constructs input and output vectors based on the original HD-sEMG signal and the decomposed MUST, trains the LSTM, and simultaneously performs polynomial regression on the discharge rate of the neural features and the torque; (5) calculates the real-time discharge rate DR of the CST through a sliding window for real-time torque estimation. The present invention is used for natural and real-time control of a robotic prosthesis, can provide a better interactive experience for people with disabilities, and can also be widely used in rehabilitation robots, human-computer interaction and other fields.

Description

Technical Field

[0001] The present invention belongs to the field of biomechanical and electrical integration, and specifically relates to an online wrist torque estimation method based on neural features and LSTM. Background Art

[0002] Electromyography (EMG) is a type of bioelectrical signal in the human body that represents the electrical activity of skeletal muscles. Skeletal movement is achieved through muscle contraction, which is always accompanied by the generation of EMG signals. Therefore, analyzing EMG signals can help us assess changes in the neuromuscular system's function. Surface EMG (sEMG) is a muscle signal measured by electrodes on the skin's surface. It is the superposition of motor unit action potential trains (MUAPs) that are low-pass filtered by a bulk conductor and then transmitted to the skin's surface. When the measuring electrodes are arranged in a high-density array, multi-channel surface EMG signals can be measured simultaneously, forming a high-density sEMG (HD-sEMG). Using a blind source separation (BSS) algorithm, HD-sEMG can be decomposed into motor unit spike trains (MUSTs), thereby revealing the deep neural signatures of muscle electrical activity.

[0003] Electromyographic signals are widely used in human-machine interaction control, including achieving natural control of operating forces and torques. Currently, there are two main approaches for torque estimation in human-machine interfaces, such as robotic arms and prosthetic limbs. One approach relies on pattern recognition based on sEMG. These methods are often limited to surface features, lack understanding of the sEMG generation mechanism, and have poor generalization and interpretability. The other approach relies on computational methods based on skeletal muscle models. These methods are complex, computationally cumbersome, and time-consuming. A torque estimation method with clear physiological significance and rapid computational time is crucial for real-time control of prosthetic hands. Summary of the Invention

[0004] To address these issues, this paper discloses an online wrist torque estimation method based on neural signatures and LSTM. This method can be used for both natural and real-time control of a robotic prosthesis, providing a better interactive experience for people with disabilities. It also has broad applications in rehabilitation robotics, human-computer interaction, and other fields. This paper provides a new control strategy that considers the neural signatures of muscle activity and constructs an online torque estimation model using deep learning methods, enabling real-time torque estimation.

[0005] In order to achieve the above-mentioned purpose, the technical solution of the present invention is as follows:

[0006] A method for online wrist torque estimation based on neural features and LSTM (Long Short-Term Memory) network includes the following steps:

[0007] (1) Training data preparation:

[0008] Before the experiment, high-density electrodes were attached to the extensor and flexor muscles of the subject's forearm. Afterward, the subject performed the designated movements according to the experimental paradigm. sEMG was collected using high-density electrodes, typically with 64, 128, or 256 channels. Torque sensors were also used to measure force. The collected data was preprocessed for subsequent offline training.

[0009] (2) Blind source separation must obtain:

[0010] A blind source separation algorithm is used to extract the preprocessed data, identifying the individual motor unit action potential train signals (MUAPt) and their corresponding separation vectors. A Kmeans clustering algorithm is then used to identify neural firing pulses, resulting in the MUST. MUST is a 0-1 sequence, where 0 indicates that the MU is not generating action potentials at the current moment, and 1 indicates that the MU is currently firing.

[0011] (3) LSTM neural network learning and polynomial regression:

[0012] A sequence-to-sequence LSTM classifier is constructed for the raw HD-sEMG and the decomposed MUSTs. Assuming the high-density electrode patch has M channels and N MUSTs are obtained through blind source separation, then at each time step, the sEMG signals from the M channels form an M-dimensional input feature vector, and the N MUST values ​​at the current time step are processed to form an N-dimensional output column vector. Using this input and output, the LSTM is trained to produce a neural network capable of identifying MU activation from raw HD-sEMG signals.

[0013] The decomposed MUSTs are accumulated to obtain a Cumulative Spike Train (CST). The number of spikes in the CST is calculated and polynomial regression is performed with the upsampled moment to determine the regression polynomial.

[0014] (4) Sliding window online estimation of torque:

[0015] For online HD-sEMG signals, each of the M channels receives a new sEMG sample at each new timestamp, forming an M-dimensional input vector. The LSTM network generates an N-dimensional output vector. A 400ms sliding window is selected, within which the accumulated pulse train is calculated, and polynomial regression is used to estimate wrist torque in real time.

[0016] Furthermore, step (1) includes:

[0017] (1.1) Filtering: For sEMG signals, a 4th-order Butterworth bandpass filter is used to filter the raw sEMG signals at 20-500 Hz, and a 50 Hz notch filter is used to remove power frequency interference in the signal. For torque signals, a sliding average filter is used to filter out noise interference.

[0018] (1.2) Centering: In Independent Component Analysis (ICA), each random variable is required to have a zero mean, and the sEMG is subtracted from its mean.

[0019] (1.3) Extension: Surface electromyography signal x i (n) can be modeled as a convolutional mixture model as follows:

[0020]

[0021] where w i (n) is random noise; The waveform of the jth MUAP in the i-th channel is represented by t, and the waveform length is L; j (n) represents the discharge moment of the jth MU, and its value is 0 or 1, which can be expressed by the unit pulse function δ: j (n)=∑ k δ(nT j (k)), that is, when the jth MU is in T j (k) activation time t j (T j (k))=1.

[0022] After the M-channel sEMG signal is expanded, the convolution mixture model of the surface electromyography can be expressed as a linear instantaneous mixture model. The observation matrix before expansion consists of M-channel sEMG, which is expressed as X = [x1(n), x2(n), ..., x M (n)] T The expanded matrix is

[0023] (1.4) Whitening: Whitening is used to achieve decorrelation of the source signal and reduce the redundancy of the data input. The covariance matrix ∑ is decomposed into eigenvalues: ∑=UΛU T , where U is an orthogonal matrix and Λ is an eigenvalue matrix, then W=UΛ -1 / 2 U T is the whitening matrix. The whitening matrix W is Multiply to get the whitened data Z.

[0024] Furthermore, step (2) includes:

[0025] (2.1) The separation vector w is obtained by the fixed point iteration method: Assuming that the contrast function that measures the sparseness of the source signal is G(x), and its first-order derivative is g(x), the iterative equation of w is: w(n) = E{Zg[w(n-1) T Z]}-Aw(n-1), where A=E{g′[w i (n-1) T Orthogonalize and normalize w(n) to ensure that the separated source signals have no repeated signals.

[0026] (2.2) Separation vector improvement: Multiply the separation vector w(n) by the whitened matrix Z to obtain the source signal, and perform binary clustering to obtain the MUST. The discreteness of the MUST is measured by the coefficient of variation of inter-spike intervals (CoVofISI). This value should be as small as possible. With this as the convergence target, the separation vector is iterated again. The iterative equation is: Where K is the number of MUST values ​​1, t k The moment when the median value of MUST is 1.

[0027] (2.3) Separation vector verification: Verify the separation vectors obtained through steps (2.1) and (2.2) to improve the reliability of the separation results. The reliability of the separation results is measured by three parameters: PNR (Pulse to Noise Ratio), SIL (Silhouette Distance), and MISI (Minimum Inter-Spike Interval). These three parameters evaluate the reliability of the separated source signal from three aspects: signal-to-noise ratio, clustering effect, and whether the neural discharge principle is met. The calculation formula is as follows:

[0028]

[0029] MISI=min(ISI)

[0030] In the above PNR calculation formula, j represents the jth MU; in the SIL calculation formula, b represents the sum of the inter-cluster distances of the clusters, a represents the sum of the intra-cluster distances of the clusters; ISI represents the inter-spike interval.

[0031] Furthermore, step (3) includes:

[0032] (3.1) Constructing the input vector: At a certain moment, the sEMG values ​​of all channels constitute the input column vector I(n) = [x1(n), x2(n), ..., x M (n)] T ∈Q M .

[0033] (3.2) Constructing the Output Vector: The values ​​of all MUSTs at the current moment form an N-dimensional column vector, whose elements are either 0 or 1. However, a 1 in the vector only indicates that the MUAP's peak is at the current moment. A MU with a value of 0 is not necessarily inactive; a MU may be active and still generate an action potential, but the action potential's peak is not at the current moment. Therefore, MUSTs must be modified to serve as the LSTM output vector. Considering the typical MUAP waveform length of 90ms, the present invention expands the 1 in MUST, assigning its values ​​within the first 45ms and last 45ms from 0 to 1. The modified MUSTs are sampled at the current moment to form an N-dimensional column vector O(n).

[0034] Due to the overlap of MUAP waveforms, there may be multiple 1s in the vector O(n), which makes the LSTM output not suitable for one-hot encoding. To achieve better recognition performance for the LSTM, Schmidt orthogonalization is used to make the distances between different o(n) equal. Assume that O(n) has p different values ​​and form them into a matrix O = [o1, ..., o p ], perform UT decomposition on O and get:

[0035]

[0036] The column vectors in the matrix U satisfy orthogonalization and unit. For the vector O(n), replace it with the corresponding η as the output vector of LSTM, denoted as

[0037] (3.3) Constructing a neural network: The neural network consists of an input layer, an LSTM layer, a Batch Normalization layer, a Dropout layer, a Fully Connected layer, a Softmax layer, and an output layer. The size of the input layer is M, and the size of the output layer is N. Data set for training.

[0038] Furthermore, step (4) includes:

[0039] (4.1) Since step (3.2) performs orthogonalization and normalization, the output vector calculated by LSTM needs to be solved in reverse according to the equation of UT decomposition, which is O(n).

[0040] (4.2) Since the 1s in MUSTs are expanded, when a series of 1s are identified, it indicates that MUAP waveforms exist at these time points. When calculating neural features, only the 1s in the middle position are needed, so the other 1s need to be reassigned to 0.

[0041] (4.3) In the sliding window, add a new column vector O(n new ), remove the old column vector O(n old ), accumulate each MUST to get CST, count the pulses of CST in the window, and input the value as the independent variable into the polynomial regression equation to get the real-time estimation of the torque.

[0042] The beneficial effects of the present invention are:

[0043] 1. In terms of wrist torque prediction, this invention directly establishes the relationship between neural signatures and torque. Compared to the time-frequency domain features of sEMG, neural signatures are more stable and have clearer meanings, making them more robust when used for torque estimation. This offers significant advantages when used for natural control of prosthetic limbs and robotic arms.

[0044] 2. The present invention takes special considerations on the output vector. By expanding 1 in MUSTs, it makes it possible to train the LSMT neural network under sparse pulse sequences. As long as the collected HD-sEMG contains waveform information, the LSTM can identify MU activation and then extract neural features. At the same time, in order to improve the classification effect of the LSTM, the present invention orthogonalizes and normalizes the column vectors of the cross-section of MUSTs so that the distances between different vectors are equal, which can meet the training requirements of the LSTM. Through special processing of the output vector, the present invention makes it possible to combine the characteristics of MUST with LSTM, thereby enabling effective estimation.

[0045] 3. During the iterative process of separation vectors, the present invention tests the separation vectors from three levels: signal-to-noise ratio, clustering effect, and physiology, to ensure that the results obtained by the blind source separation algorithm are accurate and reliable.

[0046] 4. This invention proposes a sliding window online estimation method for torque, based on LSTM and neural features called CST. To achieve online real-time estimation, computational latency must be minimal. The sliding window CST counting method proposed in this invention effectively meets this low-latency requirement. When the window length is fixed, each new HD-sEMG sample generates a new input vector, which is then fed into the trained LSTM network. Following the aforementioned steps, a predicted MUSTs column vector can be rapidly calculated. The sliding window acts like a stack, with the newly predicted MUSTs column vector entering the sliding window, displacing the earliest vector in the window. The CST neural feature is obtained by accumulating the MUSTs in the sliding window, and the torque estimate is calculated by counting the number of CST pulses within the sliding window. The entire process is computationally simple, enabling continuous online real-time estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flowchart of the online wrist torque estimation based on neural features and LSTM of the present invention;

[0048] Figure 2 It is a flow chart of the blind source separation algorithm of the present invention;

[0049] Figure 3 It is a schematic diagram of LSTM training of the present invention;

[0050] Figure 4 Schematic diagram of output vector construction of the present invention;

[0051] Figure 5 Schematic diagram of online estimation of sliding window torque according to the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0053] like Figure 1 As shown, the present invention describes an online wrist torque estimation method based on neural features and LSTM, and the specific steps include:

[0054] (1) Training data preparation:

[0055] Before the experiment, high-density electrodes were attached to the extensor and flexor muscles of the subjects' forearms. Afterward, the subjects performed the designated movements according to the experimental paradigm. After the experiment began, sEMG and torque were collected using the high-density electrodes and torque sensors, respectively. The collected data were preprocessed for subsequent offline training.

[0056] Specifically, data processing includes the following steps:

[0057] (a) Filtering: For sEMG signals, a 4th-order Butterworth bandpass filter was used to filter the raw sEMG signals at 20–500 Hz, and a 50 Hz notch filter was used to remove power frequency interference from the signals. For torque signals, a sliding average filter was used to remove noise interference.

[0058] (b) Centering: Subtract the mean of sEMG to satisfy the requirement that the mean of each random variable in ICA is 0.

[0059] (c) Extension: Surface electromyography signal x i (n) can be modeled as a convolutional mixture model as follows:

[0060]

[0061] where w i (n) is random noise; The waveform of the jth MUAP in the i-th channel is represented by t, and the waveform length is L; j (n) represents the discharge moment of the jth MU, and its value is 0 or 1, which can be expressed by the unit pulse function δ: j (n)=∑ k δ(nT j (k)), that is, when the jth MU is in T j (k) activation time t j (T j (k))=1.

[0062] The observation matrix before expansion consists of M channels of sEMG, which is expressed as X = [x1(n), x2(n), ..., x M (n)] T After the sEMG signals of M channels are expanded, the expanded matrix is

[0063] (d) Whitening: Expand the matrix The covariance matrix ∑ is decomposed into eigenvalues: ∑=UΛU T , where U is an orthogonal matrix and Λ is an eigenvalue matrix, then W=UΛ -1 / 2 U T is the whitening matrix. The whitening matrix W is Multiply to get the whitened data Z.

[0064] (2) Blind source separation must obtain:

[0065] After obtaining data Z, the blind source separation algorithm is used to decompose Z to obtain each independent MUAPt and the corresponding separation vector. The Kmeans clustering algorithm is used to perform binary clustering to identify the neural discharge pulses, thereby obtaining a 0-1 sequence MUST. The algorithm flow chart is as follows: Figure 2 The 1 in MUST indicates that the MU is in the discharge activation state at the current moment.

[0066] Specifically, step (2) includes:

[0067] (a) Obtain the separation vector w by the fixed point iteration method: Assuming that the contrast function that measures the sparseness of the source signal is G(x), and its first-order derivative is g(x), then the iterative equation of w is: w(n) = E{Zg[w(n-1) T Z]}-Aw(n-1), where A=E{g′[w i (n-1) T Orthogonalize and normalize w(n) to ensure that the separated source signals have no repeated signals.

[0068] (b) Separation vector improvement: Multiply the separation vector w(n) by the whitened matrix Z to obtain the source signal, and perform binary clustering to obtain the MUST. The discreteness of the MUST is measured by the pulse interval variation coefficient, which should be as small as possible. With this as the convergence target, the separation vector is iterated again. The iterative equation is: Where K is the number of MUST values ​​1, t k The moment when the median value of MUST is 1.

[0069] (c) Separation vector verification: Verify the separation vectors obtained through steps (2.1) and (2.2) to improve the reliability of the separation results. The reliability of the separation results is measured by three parameters: PNR, SIL, and MISI. These three parameters evaluate the reliability of the separated source signal from three aspects: signal-to-noise ratio, clustering effect, and whether the neural discharge principle is met. The calculation formula is as follows:

[0070]

[0071] MISI=min(ISI)

[0072] In the above PNR calculation formula, j represents the jth MU; in the SIL calculation formula, b represents the sum of the inter-cluster distances of the clusters, a represents the sum of the intra-cluster distances of the clusters; ISI represents the inter-spike interval.

[0073] (3) LSTM neural network learning and polynomial regression:

[0074] For the original HD-sEMG and the decomposed MUSTs, a sequence-to-sequence LSTM classifier is constructed, such as Figure 3As shown in Figure 2. Assuming the high-density electrode patch has M channels and blind source separation yields N MUSTs, then at each time step, the sEMG signals from the M channels form an M-dimensional input feature vector, and the processed values ​​of the N MUSTs at the current time step form an N-dimensional output column vector. Use this input and output to train the LSTM.

[0075] The decomposed MUSTs are accumulated to obtain a Cumulative Spike Train (CST). The number of spikes in the CST is calculated and polynomial regression is performed with the upsampled moment to determine the regression polynomial.

[0076] Specifically, the LSTM training steps include:

[0077] (a) Constructing the input vector: At a certain moment, the sEMG values ​​of all channels constitute the input column vector I(n) = [x1(n), x2(n), ..., x M (n)] T ∈Q M , as shown Figure 3 shown.

[0078] (b) Construct the output vector: Figure 4 As shown, the values ​​of all MUSTs at the current moment form an N-dimensional column vector, whose elements are either 0 or 1. Specifically, considering that the typical waveform length of a MUAP is 90ms, the present invention extends the 1 in MUST, assigning values ​​from 0 to 1 in the first 45ms and last 45ms. The modified MUSTs are then sampled at the current moment to form an N-dimensional column vector O(n).

[0079] In order to organically combine the characteristics of MUST with LSTM, the present invention makes the distances between different o(n) equal through Schmidt orthogonalization, so that LSTM can obtain better recognition performance. Assume that O(n) has p different values, and form a matrix O = [o1, ..., o p ], perform UT decomposition on O and get:

[0080]

[0081] The column vectors in the matrix U satisfy orthogonalization and unit. For the vector O(n), replace it with the corresponding η as the output vector of LSTM, denoted as

[0082] (c) Constructing a neural network: The neural network consists of an input layer, an LSTM layer, a Batch Normalization layer, a Dropout layer, a Fully Connected layer, a Softmax layer, and an output layer. Its structure is shown in the figure. Figure 3 As shown. The size of the input layer is the number of channels M, and the size of the output layer is N. Data set for training.

[0083] (4) Sliding window online estimation of torque:

[0084] The method of online estimation of moment by sliding window is as follows Figure 5 As shown in the figure, for online HD-sEMG signals, each of the M channels receives a new sEMG sample at each new timestamp, forming an M-dimensional input vector. After calculation by the LSTM network, an N-dimensional output vector is generated. A 400ms sliding window is selected, and the MUSTs are accumulated to obtain the CST. The number of CST spikes is calculated, and polynomial regression is used to estimate wrist torque in real time.

[0085] Specifically, step (4) includes:

[0086] (a) Since step (3.2) performs orthogonalization and normalization, the output vector calculated by LSTM needs to be solved inversely according to the equation of UT decomposition, which is O(n).

[0087] (b) Since the 1s in MUSTs are expanded, when a series of 1s are identified, it indicates that MUAP waveforms exist at these time points. When calculating neural features, only the 1s in the middle position are needed, so the other 1s need to be reassigned to 0.

[0088] (c) In the sliding window, add a new column vector O(n new ), remove the old column vector O(n old ), accumulate each MUST to get CST, count the pulses of CST in the window, and input the value as the independent variable into the polynomial regression equation to get the real-time estimation of the torque.

Claims

1. A method for online wrist torque estimation based on neural features and LSTM, characterized by: The following steps are involved: (1) Training data preparation: Before the experiment, high-density electrodes were attached to the extensor and flexor muscles of the subjects' forearms. Afterward, the subjects performed the designated movements according to the experimental paradigm. sEMG was collected using the high-density electrodes, with 64, 128, or 256 channels. Meanwhile, torque was measured using a torque sensor. The collected data was preprocessed for subsequent offline training. (2) Blind source separation must obtain: For the preprocessed data, the blind source separation algorithm is used to obtain the independent motor unit action potential train signal MUAPt and the corresponding separation vector, and the neural discharge pulse is identified by the Kmeans clustering algorithm to obtain MUST; MUST is a 0-1 sequence, where 0 means that the MU has no action potential at the current moment, and 1 means that the MU is in a discharge activation state at the current moment; (3) LSTM neural network learning and polynomial regression: A sequence-to-sequence LSTM classifier is constructed for the original HD-sEMG and the decomposed MUSTs. Assuming the high-density electrode patch has M channels and N MUSTs are obtained through blind source separation, then at each time step, the sEMG signals from the M channels form an M-dimensional input feature vector, and the N MUST values ​​at the current time step are processed to form an N-dimensional output column vector. Using this input and output, the LSTM is trained to obtain a neural network capable of identifying MU activation from the original HD-sEMG signal. The decomposed MUSTs are accumulated to obtain CST, the number of CST spikes is calculated, and polynomial regression is performed with the upsampled moment to determine the regression polynomial. (4) Sliding window online estimation of torque: For online HD-sEMG signals, each of the M channels can obtain a new sEMG sampling data at each new timestamp, which together form an M-dimensional input vector. After calculation by the LSTM network, an N-dimensional output vector can be obtained. A 400ms sliding window was selected, the accumulated pulse train was calculated within the window, and the wrist torque was estimated in real time by polynomial regression.

2. The online wrist torque estimation method based on neural features and LSTM according to claim 1 is characterized in that: Step (1) includes: (1.1) Filtering: For sEMG signals, a 4th-order Butterworth bandpass filter is used to filter the raw sEMG signals at 20-500 Hz, and a 50 Hz notch filter is used to remove power frequency interference in the signal; for torque signals, a sliding average filter is used to filter out noise interference; (1.2) Centering: In ICA, each random variable is required to have a zero mean, and the sEMG is subtracted from its mean; (1.3) Extension: Surface electromyography signal x i (n) is modeled as a convolutional mixture model as follows: where w i (n) is random noise; The waveform of the jth MUAP in the i-th channel is represented by t, and the waveform length is L; j (n) represents the discharge moment of the jth MU, and its value is 0 or 1, which is expressed by the unit pulse function δ: j (n)=∑ k δ(nT j (k)), that is, when the jth MU is in T j (k) activation time t j (T j (k))=1; After the M-channel sEMG signals are expanded, the convolution mixture model of the surface electromyography is expressed as a linear instantaneous mixture model; the observation matrix before expansion consists of the M-channel sEMG, which is expressed as X = [x1(n), x2(n), ..., x M (n)] T ; The expanded matrix is (1.4) Whitening: Whitening is used to achieve decorrelation of the source signal and reduce the redundancy of the data input; The covariance matrix ∑ is decomposed into eigenvalues: ∑=UΛU T , where U is an orthogonal matrix and Λ is an eigenvalue matrix, then W=UΛ -1 / 2 U T is the whitening matrix; the whitening matrix W is Multiply to get the whitened data Z.

3. The online wrist torque estimation method based on neural features and LSTM according to claim 1 is characterized in that: Step (2) includes: (2.1) The separation vector w is obtained by the fixed point iteration method: Assuming that the contrast function that measures the sparseness of the source signal is G(x), and its first-order derivative is g(x), the iterative equation of w is: w(n) = E{Zg[w(n-1) T Z]}-Aw(n-1), where A=E{g′[w i (n-1) T Z]}; orthogonalize and normalize w(n) to ensure that the separated source signals have no repeated signals; (2.2) Separation vector improvement: Multiply the separation vector w(n) by the whitened matrix Z to obtain the source signal, and perform binary clustering to obtain the MUST. The discreteness of the MUST is measured by the pulse interval variation coefficient, which should be as small as possible. The separation vector is iterated again with this as the convergence target. The iterative equation is: Where K is the number of MUST values ​​1, t k The moment when the median value of MUST is 1; (2.3) Separation vector verification: Verify the separation vector obtained through steps (2.1) and (2.2) to improve the reliability of the separation results. The reliability of the separation results is measured by three parameters: PNR, SIL, and MISI. These three parameters evaluate the reliability of the separated source signal from three aspects: signal-to-noise ratio, clustering effect, and whether the neural discharge principle is met. The calculation formula is as follows: MISI=min(ISI) In the above PNR calculation formula, J represents the jth MU; in the SIL calculation formula, b represents the sum of the inter-cluster distances of the clusters, a represents the sum of the intra-cluster distances of the clusters; and ISI represents the inter-spike interval.

4. The online wrist torque estimation method based on neural features and LSTM according to claim 1, characterized in that: Step (3) includes: (3.1) Constructing the input vector: At a certain moment, the sEMG values ​​of all channels constitute the input column vector I(n) = [x1(n), x2(n), ..., x M (n)] T ∈Q M ; (3.2) Constructing the output vector: The values ​​of all MUSTs at the current moment form an N-dimensional column vector, the elements of which are either 0 or 1. However, a 1 in the vector only represents that the MUAP's peak is at the current moment. A MU with a value of 0 is not necessarily inactive. There are cases where a MU is activated and still generates an action potential, but the action potential's peak is not at the current moment. Therefore, MUSTs need to be modified to serve as the output vector of the LSTM. Considering that the waveform length of the MUAP is 90ms, the 1 in MUST is expanded, and its values ​​in the first 45ms and the last 45ms are both assigned from 0 to 1. The modified MUSTs are sampled at the current moment to form an N-dimensional column vector O(n). Due to the overlap of MUAP waveforms, there may be multiple 1s in the vector O(n), which makes the LSTM output not satisfy the one-hot encoding. In order to make LSTM obtain better recognition performance, the distances between different o(n) are made equal through Schmidt orthogonalization. Assuming that O(n) has p different values, they are combined into a matrix O = [o1, ..., o p ], perform UT decomposition on O and get: The column vectors in the matrix U satisfy orthogonalization and unit. For the vector O(n), it is replaced by the corresponding η as the output vector of LSTM, which is recorded as (3.3) Constructing a neural network: The neural network consists of an input layer, an LSTM layer, a Batch Normalization layer, a Dropout layer, a Fully Connected layer, a Softmax layer, and an output layer. The size of the input layer is the number of channels M, and the size of the output layer is N. Data set for training.

5. The online wrist torque estimation method based on neural features and LSTM according to claim 1 is characterized in that: Step (4) includes: (4.1) Since step (3.2) performs orthogonalization and normalization, the output vector calculated by LSTM needs to be solved inversely according to the equation of UT decomposition, which is O(n); (4.2) Since the 1s in MUSTs are expanded, when a series of 1s are identified, it indicates that MUAP waveforms exist at these time points. When calculating neural features, only the 1s in the middle position are needed, so the other 1s need to be reassigned to 0; (4.3) In the sliding window, add a new column vector O(n new ), remove the old column vector O(n old ), accumulate each MUST to get CST, count the pulses of CST in the window, and input the value as the independent variable into the polynomial regression equation to get the real-time estimation of the torque.

Images