A method for motion intent recognition
By processing electromyographic (EMG) signals using low-order fuzzy models and probabilistic models, the problem of EMG signals being susceptible to noise interference is solved, improving the accuracy and robustness of motor intention recognition and achieving efficient motor intention prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-13
- Publication Date
- 2026-04-03
AI Technical Summary
Human bioelectrical signals are weak and easily interfered with by environmental and equipment noise, making it difficult to recognize movement intentions. Existing technologies are unable to effectively handle nonlinear and dynamic time-varying characteristics, resulting in insufficient recognition accuracy.
A low-order fuzzy model combined with a probabilistic model is used to process electromyographic signals through filters, extract features, and establish the expected objective function and evaluation function. The kernel function and regularization parameters are optimized using sampling with replacement, and a probabilistic classification model is constructed to improve robustness and recognition efficiency.
It improves the accuracy and robustness of motion intent recognition, effectively handles nonlinear and dynamic time-varying features, reduces the impact of noise interference, and achieves a recognition accuracy rate of over 90%.
Smart Images
Figure SMS_1 
Figure SMS_4 
Figure SMS_7
Abstract
Description
Technical Field
[0001] This application relates to the field of human signal processing and recognition, and in particular to a method for recognizing motion intentions. Background Technology
[0002] Active rehabilitation robots used in the rehabilitation training of patients with spinal cord injuries can sense movement intentions based on bioelectrical signals generated by human movement, and actively adjust output torque and joint angles, enabling patients to have basic self-care abilities. However, human bioelectrical signals are weak, there are differences in electromyographic signals between individuals, and they are easily interfered with by environmental noise, equipment noise, and other bioelectrical signals, which makes the recognition of human intentions difficult.
[0003] Based on different signal types, intention recognition methods can be divided into those based on human bioelectrical signals and those based on human-computer interaction force information. Research based on human bioelectrical signals focuses on surface electromyography (sEMG) and electroencephalography (EEG), using sensors installed at different locations on the body to detect valid information and analyzing and perceiving human movement intentions using methods such as Bayesian networks, neural networks, multilayer perceptrons, fuzzy approximation, support vector machines, and neurofuzzy logic. However, sEMG and EEG possess complex nonlinear, strongly coupled, and dynamically time-varying characteristics, and are easily affected by random interference from environmental noise, acquisition equipment, and other human bioelectrical signals. This interference can cause traditional classification models to shift, leading to failure in movement intention prediction. Therefore, there is an urgent need to establish a robust human intention recognition method capable of handling nonlinear data. Summary of the Invention
[0004] To reduce interference and improve recognition efficiency and robustness, this application provides a method for remote intention recognition.
[0005] The remote motion intent recognition method provided in this application adopts the following technical solution:
[0006] A method for recognizing remote motion intent includes the following steps performed sequentially:
[0007] S1 Data Acquisition: The sEMG signal acquisition system is used to acquire and process electromyographic signals of human movement to obtain raw electromyographic signal data;
[0008] S2 Data Processing: The original data is processed by filtering to remove interference and extract the features from the original data to obtain a data feature set.
[0009] S3 establishes the desired objective function and evaluation function, specifically including:
[0010] first step,
[0011] Establish low-order model rules, and obtain the original objective function based on the low-order model rules. The original objective function is expressed as:
[0012]
[0013] Constraints:
[0014] Among them, J(a k ,e i Let ) be the objective function, and a k and b k It is the consequent parameter of the k-th rule, γ is the regularization parameter obtained by cross-validation, and e i It is the model error, φ k (x i ) is x i The membership degree of the k-th rule, It is the projection function, x i It is a subset of the data feature set of the training data, y i It is the motion tag value at time i;
[0015] The original objective function is then transformed using the Lagrange equation.
[0016] Simultaneously, optimization conditions are established and solved under these conditions to obtain a low-order fuzzy model:
[0017]
[0018] Among them, y e It is the motion label value predicted by the model, x e It is a subset of the data feature set of the predicted data. It is a membership kernel function, g ie The membership kernel function represents the motion tag values at time i and time e.
[0019] It is the kernel function, and σ is the kernel function parameter;
[0020] Substituting the low-order fuzzy model into the original objective function yields the optimized objective function:
[0021]
[0022] in, Λ=(λ1 λ2 … λ N ) T ,
[0023] λ i They are Lagrange multipliers;
[0024] The second step,
[0025] The expectation operation is performed on both the low-order fuzzy model and the optimized objective function to obtain the final evaluation function:
[0026]
[0027] The desired objective function is obtained as follows:
[0028]
[0029] Wherein, ρ(g) ie ) is the probability density function of the membership kernel function;
[0030] It is the expectation matrix;
[0031] The data feature set is processed using a final evaluation function and an expected objective function.
[0032] Preferably, it also includes probability estimates of the S4 kernel function parameters and regularization parameters:
[0033] By using sampling with replacement, a portion of the data feature set is divided into q groups as training data, and the other portion of the data is used as prediction data.
[0034] Using the training data and training the model through the first step in S3, we obtain the optimal kernel function parameters and the optimal regularization parameters;
[0035] Constructing the probability distribution models of the optimal kernel function parameters and the optimal regularization parameters: When the number of groups is large enough, the probability distribution models of the optimal kernel function parameters and the optimal regularization parameters both approximately follow a normal distribution. Substitute the probability distribution models of the optimal kernel function parameters and the optimal regularization parameters into the final evaluation function and the expected objective function.
[0036] Preferably, it also includes the distribution estimation of the S5 membership kernel function parameters:
[0037] Substituting the probability distribution model of the kernel function parameters into the membership kernel function parameters makes the membership kernel function parameters random variables and expressed as follows:
[0038] Based on the kernel function parameters, the probability density function of the kernel function parameters can be expressed as:
[0039]
[0040] Due to the membership function φ k () and kernel function k(x) i ,x j All of them are positive definite, so when At that time, there were:
[0041]
[0042] when At that time, there were:
[0043]
[0044] Where F is the distribution function, represent The probability of;
[0045] At this point, the probability density function of the model parameters is:
[0046]
[0047] Obtain the membership kernel function parameter g ij The expected values are as follows:
[0048]
[0049] The membership kernel function parameter g ij Substituting the expected value into the final evaluation function and the expectation function, we obtain the final probability classification model, which is used to process the data feature set used as prediction data.
[0050] Preferably, the sEMG signal acquisition system includes: an sEMG signal receiver, a three-lead cable, an Arduino UNO development board, surface electromyography electrodes, a front-end analog circuit for acquisition, and a back-end digital signal filtering and processing.
[0051] Front-end analog circuit acquisition: Electromyography (EMG) signals are acquired through multiple channels, and then amplified and filtered to obtain EMG signals;
[0052] Mid-range: The SPDT switch on the sEMG signal receiver switches between EnvelopeMode and RAWMode output signals. EnvelopeMode uses envelope detection processing to obtain the dynamic EMG detection signal. RAWMode outputs the raw EMG signal. The channel gain of the EnvelopeMode and RAWMode output signals can be adjusted by the blue and white potentiometers on the sEMG signal receiver. At the same time, an external single-channel AudioOutputInterface is provided to output the raw signal. The AudioOutputInterface outputs the raw waveform of a certain channel through a jumper cap, which allows users to monitor and analyze the current EMG signal in real time with the help of external instruments.
[0053] Back-end digital signal filtering and processing: Use the Arduino UNO development board to match the output signal of the acquisition middleware.
[0054] Preferably, S2 data processing is as follows:
[0055] Filtering: A 30-300Hz bandpass filter and a 50Hz notch filter with adjusted Butterworth filter parameters are used to handle low-frequency noise interference and 50Hz power frequency interference generated by mains voltage radiation.
[0056] Feature extraction: The extracted time-domain features include: root mean square (RMS), standard deviation (SD), and mean absolute value (MAV); the frequency-domain features include: average power frequency (MPF) and median frequency (MF).
[0057] A data feature set including the above-mentioned features is obtained.
[0058] In summary, this application includes at least one of the following beneficial technical effects:
[0059] 1. In order to solve the problems of human motion intention perception based on electromyography signals, improve the recognition efficiency and robustness of the model and enable it to handle nonlinear data, this invention combines the excellent predictive ability of low-order fuzzy models with robust probabilistic models, and uses expectation operation to predict motion intention to eliminate random interference in the model and improve the model's ability to handle time-varying dynamic features and external disturbances.
[0060] 2. Because surface electromyography (EMG) signals are easily affected by noise from the acquisition equipment, the environment, and other human bioelectrical signals during the acquisition process, this noise can cause the classification model to fail, leading to inaccurate prediction of movement intent. To more realistically simulate nonlinear systems under interference environments, a sampling with replacement method is used to train the model. The resulting optimal regularization parameters and kernel function parameters are then used to optimize the model, suppressing the impact of noise on movement intent recognition and maximizing the accuracy of model predictions. Detailed Implementation
[0061] The motion intent recognition method disclosed in this application includes the following steps performed sequentially:
[0062] S1 Data Acquisition: The sEMG signal acquisition system is used to acquire and process the electromyographic signals of human movement to obtain the raw data of the electromyographic signals.
[0063] Specifically:
[0064] Electromechanical signal acquisition equipment is used, which includes an sEMG signal acquisition system, LED lights, and a host computer. The sEMG signal acquisition system includes: an sEMG signal receiver, a three-lead cable, an Arduino UNO development board, surface electromyography electrodes, front-end analog circuitry for acquisition, and back-end digital signal filtering and processing. The LED lights indicate the EMG signal intensity by flashing, and the host computer stores the electromyography signals.
[0065] Front-end analog circuit acquisition: Acquires muscle electrical signals from the human leg through 1-6 channels, and then obtains EMG signals through amplification and filtering;
[0066] Mid-range: The sEMG signal receiver's SPDT switch switches between EnvelopeMode and RAWMode output signals. EnvelopeMode uses envelope detection processing to obtain the dynamic EMG detection signal; RAWMode outputs the raw EMG signal. The channel gain of the EnvelopeMode and RAWMode output signals can be adjusted via the blue and white potentiometers on the sEMG signal receiver. Simultaneously, an external single-channel AudioOutputInterface is provided to output the raw signal. This AudioOutputInterface outputs the raw waveform of a specific channel via a jumper cap, allowing users to monitor and analyze the current EMG signal in real time with the aid of external instruments.
[0067] Back-end digital signal filtering and processing: The 6-core electromyography module matched with the Arduino UNO development board is used to collect the output signal of the middle stage and control the LED light to quickly reference the EMG signal strength from the LED light; at the same time, the host computer can be used to record and view the EMG envelope signal or raw signal of up to 6 different muscles.
[0068] S2 Data Processing: The original data is processed by filtering to remove interference and extracting features to obtain a data feature set.
[0069] Specifically:
[0070] Filtering: A 30-300Hz bandpass filter and a 50Hz notch filter with adjusted Butterworth filter parameters are used to handle low-frequency noise interference and 50Hz power frequency interference generated by mains voltage radiation.
[0071] Feature extraction: The extracted time-domain features include: root mean square (RMS), standard deviation (SD), and mean absolute value (MAV); the frequency-domain features include: average power frequency (MPF) and median frequency (MF).
[0072] A data feature set including the above-mentioned features is obtained.
[0073] S3 establishes the desired objective function and evaluation function, specifically including:
[0074] S31 establishes low-order model rules:
[0075] The low-order model rule is represented as:
[0076] R k Ifx (1) is A k1 and...and x (n) is A kn , then k = 1, 2, ..., R.
[0077] Among them, R k Represents the k-th rule, x (j) A is the j-th subset of the data feature set. kj It is a fuzzy set, a k and b k It is the consequent parameter of the k-th rule. It is the projection function, y k This is the output of the k-th fuzzy rule, where R is the number of rules;
[0078] S32 obtains the expression for the original objective function through low-order model rules as follows:
[0079]
[0080] Constraints:
[0081] Among them, J(a k ,e i Let be the objective function, γ be the regularization parameter obtained by cross-validation, and e be the regularization parameter. i It is the model error, φ k (x i ) is x i The membership degree of the k-th rule, y i It is the motion tag value at time i;
[0082] S33 Transforms the original objective function using the Lagrange equation:
[0083]
[0084] Where, λ i It is a Lagrange multiplier.
[0085] S34 establishes optimization conditions and solves the low-order fuzzy model under these conditions:
[0086] The optimization conditions and solution process are as follows:
[0087]
[0088]
[0089]
[0090]
[0091] The following linear equation is obtained:
[0092] ∑θ=η
[0093] in,
[0094]
[0095]
[0096] Λ=(λ1 λ2 … λ N ) T ,
[0097] b R×1 =(b1 b2 … b R ) T y N×1 =(y1 y2 … y N ) T
[0098] Solving the linear equation yields:
[0099] θ=(∑ T ∑) -1 ∑ T η
[0100] Substituting the training data yields a low-order fuzzy model:
[0101]
[0102] Among them, y e It is the motion label value predicted by the model, x e These are the signal characteristics of the predicted data. It is a membership kernel function, g ie Let represent the membership kernel function formed by the i-th sample and the e-th sample. It is the kernel function, and σ is the kernel function parameter;
[0103] S35 optimizes the original objective function:
[0104] Substituting the solution obtained under the optimization conditions in S34 into the expression of the original objective function, we obtain the optimized objective function:
[0105]
[0106] in, Λ=(λ1 λ2 … λ N ) T ,
[0107] S36:
[0108] The final evaluation function obtained by performing expectation operation on the low-order fuzzy model is:
[0109]
[0110] Wherein, ρ(g) ie ) is the probability density function of the membership kernel function;
[0111] The desired objective function is obtained by performing expectation calculation on the optimized objective function in S35:
[0112]
[0113] in, It is the expectation matrix.
[0114] Because the regularization parameter γ in the objective function after Lagrange transformation, and the kernel function parameter σ in the low-order fuzzy model are obtained by cross-validation and are highly dependent on the data, the robustness of the model will be insufficient when the data features are disturbed. In order to eliminate the interference, S4 is used.
[0115] Probabilistic estimation of S4 kernel function parameters and regularization parameters:
[0116] S41: Perform probability estimation on the kernel function parameters, specifically:
[0117] Sampling with replacement is used to divide a portion of the data feature set into q groups as training data, and the other portion of the data feature set as prediction data.
[0118] The model was trained using training data through S31-S34, and the kernel function parameters with the best performance were selected as the optimal kernel function parameters.
[0119] Constructing a probability distribution model for the optimal kernel function parameters: When the number of groups is large enough, the probability distribution model for the optimal kernel function parameters approximately follows a normal distribution.
[0120] σ~N(μ,θ 2 ),
[0121] Where μ is the center of the normal distribution, and θ is the variance of the normal distribution; obtained by the maximum likelihood method:
[0122]
[0123]
[0124] S42: Perform probability estimation on the regularization parameter, specifically:
[0125] Sampling with replacement is used to divide a portion of the data feature set into q groups as training data, and the other portion of the data feature set as prediction data.
[0126] The optimal regularization parameters are obtained by training the model using the training data through S31-S34.
[0127] Constructing a probability distribution model for the optimal regularization parameter: When the number of groups is large enough, the probability distribution model for the optimal regularization parameter approximately follows a normal distribution.
[0128] γ~N(χ,ξ 2 ),
[0129] Where μ is the center of the normal distribution, and θ is the variance of the normal distribution; obtained by the maximum likelihood method:
[0130]
[0131]
[0132] S44: Substitute the probability distribution models of the optimal kernel function parameters and the optimal regularization parameters into the final evaluation function and the expected objective function.
[0133] Since the parameters of the kernel function approximately follow a normal distribution, the membership kernel function also becomes a random variable, expressed as:
[0134] S5 derives the distribution law of variable membership kernel functions by analyzing the distribution of kernel function parameters:
[0135] Based on the kernel function parameters, the probability density function of the kernel function parameters can be expressed as:
[0136]
[0137] Due to the membership function φ k () and kernel function k(x) i ,x j All of them are positive definite, so when At that time, there were:
[0138]
[0139] when At that time, there were:
[0140]
[0141]
[0142] Here, F is the distribution function, and the derivative of F is the probability density function. represent The probability of;
[0143] At this point, the probability density function of the model parameters is:
[0144]
[0145] Obtain the membership kernel function parameter g ij The expected values are as follows:
[0146]
[0147] Substituting the expectation of the kernel function into the final evaluation function and the expected objective function in S36, we obtain the final probabilistic classification model. The probabilistic classification model is used to process the data feature set used as prediction data.
[0148] Experimental verification:
[0149] An experiment was conducted using one of the above-mentioned methods for recognizing motor intentions and an electromyography (EMG) signal acquisition device for its application.
[0150] Four muscles were selected as EMG signal sources: rectus femoris, biceps femoris, tibialis anterior, and gastrocnemius.
[0151] Two men, aged 26 and 28, were selected and performed gait data acquisition using an sEMG system. The data included standing, feet off the ground, feet touching the ground, climbing stairs, and descending stairs, with each movement repeated 10 times for 5 seconds. After acquiring the raw signals, Butterworth filtering was applied. A sixth-order 30-350Hz bandpass filter and a 50Hz notch filter were used to handle noise interference.
[0152] Feature selection: Three feature values—variance (VAR), root mean square (RMS), and average power frequency (MPF)—were selected as the data feature set for intention recognition. In this experiment, a moving data window was used to process the active segment information of each cycle of surface electromyography (EMG) signals, with a moving data window length of 100 ms and a moving distance of 100 ms for each moving data window.
[0153] During training, 50% of the data feature set is divided into 40 groups, each containing 500 data points. For each group, steps S31-S34 are used to train the model to obtain the optimal kernel function parameters and optimal regularization parameters. Then, steps S35-S5 are used to construct a probabilistic classification model. The remaining 50% of the data is used to identify motion intentions through the final probabilistic classification model.
[0154] Motion intent recognition results:
[0155] Table 1. Accuracy of Motion Intent Recognition
[0156]
[0157] As shown in Table 1, compared with other methods or models, the action recognition accuracy obtained by using this invention is greater than 90%. In particular, this experiment selected actions with high similarity, which can greatly interfere with the accuracy of motion intention recognition. However, this invention still has a higher prediction accuracy compared with existing methods or models.
[0158] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. A method for recognizing motion intent, characterized in that: This includes the following steps performed sequentially: S1 Data Acquisition: The sEMG signal acquisition system is used to acquire and process electromyographic signals of human movement to obtain raw electromyographic signal data; S2 Data Processing: The original data is processed by filtering to remove interference and extract the features from the original data to obtain a data feature set. S3 establishes the desired objective function and evaluation function, specifically including: first step, Establish low-order model rules, and obtain the original objective function based on the low-order model rules. The original objective function is expressed as: ， Constraints: ; in, Let be the objective function. and It is the first The consequent parameter of the rule, It is a regularization parameter, obtained by cross-validation. It is model error. yes For the The membership degree of a rule. It is a projection function. It is a subset of the data feature set of the training data. It is the first The motion tag value at each moment; The original objective function is then transformed using the Lagrange equation. Simultaneously, optimization conditions are established and solved under these conditions to obtain a low-order fuzzy model: ， in, These are the motion label values predicted by the model. It is a subset of the data feature set of the prediction data. It is a membership kernel function. Representing the The moment and the The membership kernel function composed of the motion tag values at each time step. It's a kernel function. These are kernel function parameters; Substituting the low-order fuzzy model into the original objective function yields the optimized objective function: ， in, , , , i They are Lagrange multipliers; The second step, The expectation operation is performed on both the low-order fuzzy model and the optimized objective function to obtain the final evaluation function: ， The desired objective function is obtained as follows: ， in, The probability density function of the membership kernel function; It is the expectation matrix; The data feature set is processed using a final evaluation function and an expected objective function.
2. The motion intent recognition method according to claim 1, characterized in that: It also includes probability estimates for the S4 kernel function parameters and regularization parameters: By using sampling with replacement, a portion of the data feature set is divided into q groups as training data, and the other portion of the data is used as prediction data. Using the training data and training the model through the first step in S3, we obtain the optimal kernel function parameters and the optimal regularization parameters; Constructing the probability distribution models of the optimal kernel function parameters and the optimal regularization parameters: When the number of groups is large enough, the probability distribution models of the optimal kernel function parameters and the optimal regularization parameters both approximately follow a normal distribution. Substitute the probability distribution models of the optimal kernel function parameters and the optimal regularization parameters into the final evaluation function and the expected objective function.
3. The motion intent recognition method according to claim 2, characterized in that: It also includes the distribution estimation of the S5 membership kernel function parameters: Substituting the probability distribution model of the kernel function parameters into the membership kernel function parameters makes the membership kernel function parameters random variables and expressed as follows: ; Based on the kernel function parameters, the probability density function of the kernel function parameters can be expressed as: ； Due to the membership function and kernel function All are positive definite, so when At that time, there were: ； when At that time, there were: ， ； Where F is the distribution function, represent The probability of; At this point, the probability density function of the model parameters is: ； Obtain the membership kernel function parameters The expected values are as follows: ； Membership kernel function parameters Substituting the expected value into the final evaluation function and the expected objective function, we obtain the final probabilistic classification model, which is used to process the data feature set used as prediction data.
4. The motion intent recognition method according to claim 1, characterized in that: The sEMG signal acquisition system includes: an sEMG signal receiver, a three-lead cable, an Arduino UNO development board, surface electromyography electrodes, front-end analog circuitry for acquisition, and back-end digital signal filtering and processing. Front-end analog circuit acquisition: Electromyography (EMG) signals are acquired through multiple channels, and then amplified and filtered to obtain EMG signals; Mid-range: The SPDT switch on the sEMG signal receiver switches between EnvelopeMode and RAWMode output signals; EnvelopeMode uses envelope detection processing to obtain the EMG dynamic detection signal; RAWMode outputs the raw EMG signal, and the channel gain of the EnvelopeMode and RAWMode output signals can be adjusted by the blue and white potentiometers on the sEMG signal receiver; at the same time, an external single-channel AudioOutputInterface is provided to output the raw signal, and the AudioOutputInterface outputs the raw waveform of a certain channel through a jumper cap; Back-end digital signal filtering and processing: Use the Arduino UNO development board to match the output signal of the acquisition middleware.
5. The motion intent recognition method according to claim 1, characterized in that: S2 data processing, specifically: Filtering: A 30-300Hz bandpass filter and a 50Hz notch filter with adjusted Butterworth filter parameters are used to handle low-frequency noise interference and 50Hz power frequency interference generated by mains voltage radiation. Feature extraction: The extracted time-domain features include: root mean square (RMS), standard deviation (SD), and mean absolute value (MAV); the frequency-domain features include: mean power frequency (MPF) and median frequency (MF). A data feature set including time-domain and frequency-domain features is obtained.