An adaptive functional electrical stimulation method based on motion intention recognition

Abstract

The application discloses a self-adaptive functional electrical stimulation method based on motion intention recognition, comprising the following steps: firstly, collecting surface electromyogram signals and inertial motion unit (IMU) signals of a lower limb of a human body synchronously through a wearable device; realizing data synchronization on the collected two different sampling rate heterogeneous signals; inputting the synchronized multi-modal feature data into a particle swarm optimization (PSO)-long short-term memory attention (LSTM) network (PSO-LSTMAT) model, and performing real-time STS cycle six-stage posture recognition; according to the recognized specific motion stage, applying self-adaptive functional electrical stimulation to a target muscle group, and forming a closed-loop auxiliary process. Through the introduction of the PSO-LSTMAT hybrid neural network, when processing multi-sensor combined signals, the application can effectively realize cross-modal time alignment, significantly improve the recognition accuracy of each stage in the STS motion, and avoid the recognition error of the traditional single sensor method in the action transition stage.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, and in particular relates to an adaptive functional electrical stimulation method based on motion intention recognition. Background Technology

[0002] Functional electrical stimulation (FES) technology is widely used in assisted strength training. FES can effectively improve muscle strength, motor function, and stability by applying electrical stimulation to target muscle groups at specific times. Currently, most existing FES systems rely on surface electromyography (sEMG) signals, inertial measurement unit (IMU) signals, or a combination of both to trigger electrical stimulation.

[0003] Recent research has focused on "gait phase detection + adaptive stimulation parameters." For example, stimulation is triggered based on gait phase (swing initiation and cessation, heel strike, etc.), and stimulation parameters are set using rules or threshold methods. However, this is prone to inaccuracies under complex transitions and individual differences (US10773079B2). Another example is wearable multi-sensor systems that integrate IMU and multiple EMG channels in the forearm for closed-loop neuroelectromyography stimulation and interactive control, emphasizing multimodal input but offering limited description of temporal alignment and stage-level recognition under asynchronous sampling (US11493993B2). There are also improvement schemes for FES hardware architecture and multiple surface stimulation, focusing on channel multiplexing and system structure optimization, but these do not address the issues of motion recognition granularity and trigger latency (US11247048B2).

[0004] Domestic research has introduced IMU and visual / video streams into multimodal fusion and exoskeleton control, but it mainly serves motion intention / gait prediction and does not fully cover the phase-specific triggering and timing alignment issues of FES (CN115592690A). There are also patents for human-computer interaction / robot control that integrate EMG and IMU, which provide a reference for multimodal recognition, but do not address the mapping of electrical stimulation parameters for FES closed-loop control (CN112405539B).

[0005] In summary, existing solutions either rely on a single threshold / rule, resulting in limited granularity for stage identification; or, while possessing multimodal acquisition capabilities, they lack heterogeneous sampling rate alignment, conflict modeling, and phase-specific adaptive mapping. Summary of the Invention

[0006] This invention fills the gap mentioned above by using a unified time axis fused with sEMG-IMU, conflict quantification index and PSO-LSTM-Attention (PSO-LSTMAT) stage recognition to achieve high-precision multi-stage recognition and phase-specific adaptive stimulation.

[0007] This invention discloses an adaptive functional electrical stimulation method based on motion intention recognition, which includes the following steps:

[0008] Step S1: First, the surface electromyography (sEMG) signal and the inertial motion unit (IMU) signal of the human lower limb are simultaneously collected using a wearable device;

[0009] Step S2: Perform time-domain alignment processing based on biomechanical characteristic peaks on the two heterogeneous signals with different sampling rates to achieve data synchronization;

[0010] Step S2 includes:

[0011] Step S21: Extract all extreme points of the IMU signal and sEMG signal in the time series to form a set of time-amplitude pairs;

[0012] Step S22: Utilizing the inherent biomechanical coupling relationship between the peak values ​​of sEMG signals corresponding to muscle activation and the peak values ​​of IMU signals corresponding to changes in limb angular velocity during STS exercise, the peak values ​​of IMU signals are correlated with the odd-order peak values ​​of sEMG signals, and the valley values ​​of IMU signals are correlated with the even-order peak values ​​of sEMG signals.

[0013] Step S23: After establishing the peak correspondence, the low sampling rate signal is linearly interpolated between two adjacent peak points using the time axis of the high sampling rate signal as a reference, thereby generating a new signal sequence that is completely aligned with the high sampling rate signal at the time point, and finally obtaining the synchronized surface electromyography (sEMG) signal and inertial motion unit (IMU) signal feature set.

[0014] Step S3: Input the synchronized multimodal feature data into a particle swarm optimization-based attention long short-term memory network (PSO-LSTMAT) model to perform real-time STS cycle six-stage pose recognition.

[0015] In step S4, the system's main controller, based on the identified specific movement stage, applies adaptive functional electrical stimulation (aFES) to the target muscle group through a dual-channel stimulation unit. This stimulation is precisely synchronized with the movement stage and dynamically changes in intensity, forming a closed-loop auxiliary process. After completing one STS cycle, this process is repeated.

[0016] Furthermore, in step S21, the set of time-amplitude pairs formed by all extreme points of the IMU signal and sEMG signal is as follows:

[0017] ;

[0018] For IMU signals, For sEMG signals, Let i be the IMU signal at time point i. The IMU signal at point i, Let i be the time point of the sEMG signal. The sEMG signal at point i.

[0019] Furthermore, in step S22, the peak value of the IMU signal corresponds to the odd-order peak value of the sEMG signal, and the valley value of the IMU signal corresponds to the even-order peak value of the sEMG signal as follows:

[0020] ;

[0021] ;

[0022] Let i be the time point at the i-th peak of the IMU signal. The i-th peak value of the IMU signal. Let i be the time point at the i-th odd-numbered peak of the sEMG signal. The i-th odd-numbered peak value of the sEMG signal. Let i be the time point of the IMU signal at the i-th valley point. Let i be the i-th valley point of the IMU signal. Let i be the time point at the i-th even-numbered peak of the sEMG signal. It represents the i-th even-numbered peak value of the sEMG signal.

[0023] Furthermore, in step S23, the synchronized inertial motion unit (IMU) signal and surface electromyography (sEMG) signal feature set... and Its interpolation formula is:

[0024] ;

[0025] The synchronized feature set is as follows:

[0026] ;

[0027] For the IMU signal at point m after time-domain synchronization, For time point m after time-domain synchronization, The time point corresponding to the IMU signal at point i+1. The IMU signal at point i+1 This is the feature set of the IMU after time-domain synchronization. This is the sEMG feature set after time-domain synchronization. This refers to the sEMG signal at point m after time-domain synchronization.

[0028] Furthermore, step S3 also includes the following steps:

[0029] Step S31: Construct a hybrid neural network model;

[0030] Step S32: Use the particle swarm optimization algorithm to optimize hyperparameters;

[0031] The Particle Swarm Optimization (PSO) algorithm globally optimizes the network's key hyperparameters to find the optimal combination of model parameters, thereby maximizing the accuracy of pose recognition. Key hyperparameters include the number of LSTM units, Dropout rate, learning rate, and the number of fully connected layer units. The particle velocity and position update formulas are as follows:

[0032]

[0033]

[0034] in, For inertial weights, Let be the velocity of the particle in the d-th iteration. Let be the velocity of the particle in the (d-1)th iteration. For individual learning factors, For the random number set 1, As a group learning factor, For the random number set 2, Let be the position of the particle in the d-th iteration. Let be the position of the particle in the (d+1)th iteration. This is the optimal position in the particle's history. This is the globally optimal position for the population.

[0035] Furthermore, step S4 also includes the following steps:

[0036] Step S41: Establish a phase-muscle-stimulus correspondence strategy;

[0037] Step S42: Implement a stimulus program with dynamic intensity adaptation.

[0038] Furthermore, in step S41, the STS cycle is precisely divided into six phases, and target muscle groups are set for different phases; in the "standing start" and "standing momentum transfer" phases, the quadriceps are activated through channel one to assist knee extension; in the "sitting start" and "sitting momentum absorption" phases, the hamstrings are activated through channel two to assist controlled knee flexion.

[0039] Furthermore, in step S42, the intensity of the electrical stimulation is not constant but dynamically changes within each motor phase;

[0040] In the early stages of muscle activation, apply a low-intensity electrical current to reduce muscle fatigue;

[0041] During critical periods of momentum transfer or absorption that require maximum muscle strength, the stimulation intensity is increased to 70% of the user's maximum tolerance RMS, thereby achieving the optimal strength assistance effect while ensuring safety.

[0042] This stimulation uses a biphasic wave with a frequency of 40Hz and a pulse width of 200µs to ensure muscle activation efficiency and reduce the risk of skin irritation. The beneficial effects achieved by this invention are:

[0043] Improved recognition accuracy: By introducing the PSO-LSTMAT hybrid neural network, this invention can effectively achieve cross-modal temporal alignment when processing multi-sensor (sEMG and IMU) fused signals, and significantly improve the recognition accuracy of each stage in the "sit-stand-sit" (STS) motion, avoiding the recognition error that occurs in the transition stage of the action in the traditional single sensor method.

[0044] Real-time closed-loop control: After identifying a specific movement stage, the device of the present invention can immediately drive the electrical stimulation unit to perform electrical stimulation on the target muscle group in precise synchronization with the stage movement, realizing real-time dynamic adjustment of electrical stimulation parameters, overcoming the limitations of existing FES devices that rely on fixed timing or static threshold triggering.

[0045] Personalized Adaptability: The device can automatically optimize stimulation parameters based on the user's individual differences (such as tolerance and muscle response sensitivity), so that the stimulation intensity can gradually change with the exercise stage, thereby meeting the needs of strength assistance while reducing the risk of muscle fatigue.

[0046] Safety and comfort: This invention uses a symmetrical biphasic wave electrical stimulation method and sets up a current upper limit protection mechanism to effectively avoid tissue polarization and skin discomfort caused by monophasic stimulation, thereby improving the safety of the device in assisted training and user compliance. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the present invention.

[0048] Figure 2 Temporal alignment processing for multi-sensor signals;

[0049] Figure 3 This describes the real-time attitude recognition process based on the PSO-LSTMAT model. Detailed Implementation

[0050] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0051] like Figure 1 As shown, this invention provides an adaptive functional electrical stimulation method based on motion intention recognition, which includes the following steps:

[0052] Step S1: First, the surface electromyography (sEMG) signal and inertial motion unit (IMU) signal of the human lower limb are collected synchronously through wearable devices;

[0053] Step S2: Perform time-domain alignment processing based on biomechanical characteristic peaks on the two heterogeneous signals with different sampling rates to achieve data synchronization;

[0054] Step S3: Input the synchronized multimodal feature data into a particle swarm optimization-based attention long short-term memory network (PSO-LSTMAT) model to perform real-time STS cycle six-stage pose recognition.

[0055] In step S4, the system's main controller, based on the identified specific movement stage, applies adaptive functional electrical stimulation (aFES) to the target muscle group (quadriceps or hamstrings) through a dual-channel stimulation unit. This stimulation is precisely synchronized with the movement stage and dynamically changes in intensity, forming a closed-loop auxiliary process. After completing one STS cycle, this process is repeated.

[0056] like Figure 2 As shown, in step S2, the time-domain alignment process of the multi-sensor signals further includes the following steps:

[0057] Since the sampling frequencies of sEMG (sampling rate 1500Hz) and IMU (sampling rate 96Hz) are different, this invention adopts a peak alignment algorithm based on biomechanical coupling to solve the data mismatch problem.

[0058] Step S21, extract the extreme point pairs of the feature signal:

[0059] For IMU signals (denoted as) ) and sEMG signal (denoted as Extract all extreme points (including peaks and valleys) of the time series to form a set of time-amplitude pairs.

[0060] The formula is:

[0061]

[0062] Step S22, establish cross-sensor peak correspondence:

[0063] By utilizing the biomechanical coupling relationship between muscle activation (sEMG signal peak) and limb angular velocity change (IMU signal peak) during STS exercise, the peak value of the sEMG signal is correlated with the odd-order peak value of the IMU signal, and the valley value of the sEMG signal is correlated with the even-order peak value of the IMU signal.

[0064] The correspondence is as follows:

[0065]

[0066]

[0067] Step S23, linear interpolation to achieve signal synchronization:

[0068] After establishing the peak correspondence, using the time axis of the high sampling rate signal as a reference, linear interpolation is performed on the low sampling rate signal between two adjacent peak points, thereby generating a new signal sequence that is perfectly aligned with the high sampling rate signal in terms of time points, ultimately yielding the synchronized feature set. and .

[0069] Its interpolation formula is:

[0070]

[0071] The synchronized feature set is as follows:

[0072]

[0073] like Figure 3 As shown, in step S3, the real-time pose recognition process based on the PSO-LSTMAT model also includes the following steps:

[0074] Step S31, construct a hybrid neural network model:

[0075] This invention constructs a hybrid neural network combining a Long Short-Term Memory (LSTM) network and a temporal attention mechanism. The LSTM network is used to capture long-term dependencies in synchronized time-series data from sEMG and IMU; the temporal attention mechanism is used to dynamically weight the most critical time interval features of pose transitions, amplifying discriminative information and improving the model's accuracy in recognizing motion transition phases.

[0076] The neural network model LSTM-Attention, which obtains the optimal hyperparameters based on the PSO algorithm, is referred to as the PSO-LSTMAT model in this paper. The model consists of an input layer, an LSTM layer, a Dropout layer, an Attention layer, a pooling layer, a fully connected layer, and an output layer.

[0077] The input layer takes the temporally synchronized multi-sensor dataset obtained in step S2 as input and extracts its temporal features from both. In the PSO-LSTM-Attention model, and The temporal feature matrix X obtained after the "sliding window" is then input into the neural network model.

[0078] Specifically, for We use a sliding window with a length of 200 and a step size of 50 to calculate classical time-domain indices within the window:

[0079]

[0080]

[0081]

[0082]

[0083] Then combine them into a 4-dimensional row vector, denoted as .

[0084]

[0085] Similarly, for Also, extract the time-domain metrics of acceleration and angular velocity within the synchronization window:

[0086]

[0087]

[0088]

[0089]

[0090] Concatenate them into a 4-dimensional row vector, denoted as

[0091]

[0092] The two vectors are concatenated column-wise and then normalized using Z-score to form the feature matrix X.

[0093]

[0094] The feature matrix X is read by LSTM to read the temporal context relationship. The Dropout layer prevents overfitting. The Attention layer weights the importance of key frames in the temporal sequence. After passing through the pooling layer, fully connected layer and output layer, it outputs 6 label numbers 0, 1, 2, 3, 4 and 5, which correspond to the six stages of STS (standing start, standing momentum transfer, stable standing, sitting start, sitting momentum absorption and stable sitting posture).

[0095] Step S32: Use the Particle Swarm Optimization (PSO) algorithm to optimize hyperparameters.

[0096] To improve the performance of hybrid neural networks, this invention introduces a particle swarm optimization algorithm to globally optimize key hyperparameters of the network (including the number of LSTM units, Dropout rate, learning rate, and number of fully connected layer units) to find the optimal combination of model parameters, thereby maximizing the accuracy of pose recognition. The particle velocity and position update formulas are as follows:

[0097]

[0098]

[0099] in, and Let be the velocity and position of the particle in the d-th iteration, respectively. This is the optimal position in the particle's history. This is the globally optimal position for the population.

[0100] This model can achieve high-precision real-time recognition of the six stages of STS (standing start, standing momentum transfer, stable standing, sitting start, sitting momentum absorption, and stable sitting posture).

[0101] In step S4, the closed-loop adaptive functional electrical stimulation procedure further includes the following steps:

[0102] Step S41, establish a phase-muscle-stimulus correspondence strategy:

[0103] This invention employs dual-channel stimulation, precisely dividing the STS cycle into six phases and setting target muscle groups for each phase. In the "standing initiation" and "standing momentum transfer" phases, channel one activates the quadriceps to assist knee extension; in the "sitting initiation" and "sitting momentum absorption" phases, channel two activates the hamstrings to assist controlled knee flexion; in the "stable standing" and "stable sitting" phases, channels one and two are deactivated to reduce muscle fatigue caused by electrical stimulation.

[0104] Step S42, implement the intensity-adaptive stimulation protocol:

[0105] Within each phase of movement, the intensity of electrical stimulation is not constant but dynamically varied. In the early stages of muscle activation, a lower current intensity is applied to reduce muscle fatigue; during critical periods requiring maximum muscle force for momentum transfer or absorption, the stimulation intensity is increased to 70% of the user's maximum tolerance RMS, thus achieving optimal strength assistance while ensuring safety. This stimulation uses a biphasic wave with a 40Hz frequency and a 200µs pulse width to ensure efficient muscle activation and reduce the risk of skin irritation.

[0106] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the scope of protection of the present invention; all technical solutions formed by equivalent transformations or equivalent substitutions fall within the scope of protection of the present invention; the parts of the present invention not described in detail are well-known technologies to those skilled in the art.

Claims

1. An adaptive functional electrical stimulation method based on motion intention recognition, characterized in that, The adaptive functional electrical stimulation method based on motion intention recognition is a non-therapeutic method, comprising the following steps: Step S1: First, the surface electromyography (sEMG) signal and the inertial motion unit (IMU) signal of the human lower limb are simultaneously collected using a wearable device; Step S2: Perform time-domain alignment processing based on biomechanical characteristic peaks on the two heterogeneous signals with different sampling rates to achieve data synchronization; Step S2 includes: Step S21: Extract all extreme points of the IMU signal and sEMG signal in the time series to form a set of time-amplitude pairs; Step S22: Utilizing the intrinsic biomechanical coupling relationship between the peak values ​​of sEMG signals corresponding to muscle activation during STS exercise and the peak values ​​of IMU signals corresponding to changes in limb angular velocity, the peak values ​​of IMU signals are correlated with the odd-order peak values ​​of sEMG signals, and the valley values ​​of IMU signals are correlated with the even-order peak values ​​of sEMG signals. Step S23: After establishing the peak correspondence, the low sampling rate signal is linearly interpolated between two adjacent peak points using the time axis of the high sampling rate signal as a reference, thereby generating a new signal sequence that is completely aligned with the high sampling rate signal at the time point, and finally obtaining the synchronized surface electromyography (sEMG) signal and inertial motion unit (IMU) signal feature set. Step S3: Input the synchronized multimodal feature data into a particle swarm optimization-based attention long short-term memory network (PSO-LSTMAT) model to perform real-time STS cycle six-stage pose recognition. In step S4, the system's main controller, based on the identified specific movement stage, applies adaptive functional electrical stimulation (aFES) to the target muscle group through a dual-channel stimulation unit. This stimulation is precisely synchronized with the movement stage and dynamically changes in intensity, forming a closed-loop auxiliary process. After completing one STS cycle, this process is repeated.

2. The adaptive functional electrical stimulation method based on motion intention recognition according to claim 1, characterized in that, In step S21, the set of time-amplitude pairs for all extreme points of the IMU signal and sEMG signal is as follows: ； For IMU signals, It is an sEMG signal. Let i be the IMU signal at time point i. The IMU signal at point i, Let i be the time point of the sEMG signal. The sEMG signal is at point i.

3. The adaptive functional electrical stimulation method based on motion intention recognition according to claim 1, characterized in that, In step S22, the peak value of the IMU signal corresponds to the odd-order peak value of the sEMG signal, and the valley value of the IMU signal corresponds to the even-order peak value of the sEMG signal as follows: ； ； Let i be the time point at the i-th peak of the IMU signal. The i-th peak value of the IMU signal. Let i be the time point at the i-th odd-numbered peak of the sEMG signal. The i-th odd-numbered peak value of the sEMG signal. Let i be the time point of the IMU signal at the i-th valley point. The i-th valley point of the IMU signal. Let i be the time point at the i-th even-numbered peak of the sEMG signal. It represents the i-th even-numbered peak value of the sEMG signal.

4. The adaptive functional electrical stimulation method based on motion intention recognition according to claim 1, characterized in that, In step S23, the synchronized inertial motion unit (IMU) signal and surface electromyography (sEMG) signal feature set are... and Its interpolation formula is: ； The synchronized feature set is as follows: ； For the IMU signal at point m after time-domain synchronization, For time point m after time-domain synchronization, The time point corresponding to the IMU signal at point i+1. The IMU signal at point i+1 This is the feature set of the IMU after time-domain synchronization. This is the sEMG feature set after time-domain synchronization. This refers to the sEMG signal at point m after time-domain synchronization.

5. The adaptive functional electrical stimulation method based on motion intention recognition according to claim 1, characterized in that, Step S3 also includes the following steps: Step S31: Construct a hybrid neural network model; Step S32: Use the particle swarm optimization algorithm to optimize hyperparameters; The Particle Swarm Optimization (PSO) algorithm globally optimizes the network's key hyperparameters to find the optimal combination of model parameters, thereby maximizing the accuracy of pose recognition. Key hyperparameters include the number of LSTM units, Dropout rate, learning rate, and the number of fully connected layer units. The particle velocity and position update formulas are as follows: ； ； in, For inertial weights, Let be the velocity of the particle in the d-th iteration. Let be the velocity of the particle in the (d-1)th iteration. For individual learning factors, For the random number set 1, As a group learning factor, For the random number set 2, Let be the position of the particle in the d-th iteration. Let be the position of the particle in the (d+1)th iteration. This is the optimal position in the particle's history. This is the globally optimal position for the population.

6. The adaptive functional electrical stimulation method based on motion intention recognition according to claim 1, characterized in that, Step S4 also includes the following steps: Step S41: Establish a phase-muscle-stimulus correspondence strategy; Step S42: Implement a stimulus program with dynamic intensity adaptation.

7. The adaptive functional electrical stimulation method based on motion intention recognition according to claim 6, characterized in that, In step S41, the STS cycle is precisely divided into six phases, and target muscle groups are set for different phases. In the "standing start" and "standing momentum transfer" phases, the quadriceps are activated through channel one to assist knee extension. In the "sitting start" and "sitting momentum absorption" phases, the hamstrings are activated through channel two to assist controlled knee flexion.

8. The adaptive functional electrical stimulation method based on motion intention recognition according to claim 6, characterized in that, In step S42, the intensity of the electrical stimulation is not constant but dynamically changes within each motor phase; In the early stages of muscle activation, apply a low-intensity electrical current to reduce muscle fatigue; During critical periods of momentum transfer or absorption that require maximum muscle strength, the stimulation intensity is increased to 70% of the user's maximum tolerance RMS, thereby achieving the optimal strength assistance effect while ensuring safety. The stimulation uses a biphasic wave with a frequency of 40 Hz and a pulse width of 200 µs to ensure muscle activation efficiency and reduce the risk of skin irritation.

Images