Exoskeleton driving method based on real moment constraint and continuous gait phase modeling

Abstract

The invention discloses an exoskeleton driving method based on real moment constraint and continuous gait phase modeling, which comprises the following steps: acquiring multi-modal signal data in the gait process of a subject by using a multi-sensor synchronous acquisition system, including IMU signals, EMG signals, plantar pressure signals and optical motion capture data; preprocessing the multi-mode signal data to obtain a preprocessed signal; performing continuous gait phase prediction and gait pattern recognition according to the preprocessed signal; estimating a joint instantaneous moment by using a pre-trained moment estimation network based on the predicted gait phase and gait mode; and converting the obtained instantaneous moment of the joint into a driving signal of an exoskeleton actuator. According to the method, discrete event triggering is replaced by continuous gait phase modeling, so that the smoothness and robustness of control are ensured; meanwhile, a mechanical real constraint and individualized moment estimation mechanism is introduced, stable man-machine cooperative control is achieved, and the adaptability and use experience of the exoskeleton in a complex environment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent exoskeleton control, and in particular relates to an exoskeleton driving method based on real torque constraints and continuous gait phase modeling. Background Technology

[0002] Exoskeletons, as a wearable robotic technology, have gained increasing attention in recent years in fields such as rehabilitation medicine, elderly assistance, exercise enhancement, and industrial load reduction. By providing assistive torque to joints such as the hip and knee, exoskeletons can reduce the user's muscle burden, improve mobility, and help people with mobility impairments regain walking function. However, achieving stable and natural human-machine collaborative control in complex environments remains a core technical challenge in this field.

[0003] Currently, traditional exoskeleton control methods mainly suffer from the following problems: First, control methods based on discrete gait events are ill-suited for complex scenarios. Traditional methods often use discrete events such as heel strike and toe lift as trigger conditions to update control parameters. While this may achieve basic gait following under ideal conditions, it is highly susceptible to misjudgment or delay in environments such as variable speed walking, uphill / downhill walking, and turning. This can cause abrupt changes in control signals, affecting comfort and potentially posing safety hazards. More importantly, discrete event descriptions cannot reflect the continuity and periodicity of gait, limiting the smoothness and robustness of the control strategy.

[0004] Second, torque-driven methods lack realistic constraints and smooth mapping. Existing exoskeletons often use preset trajectories or fixed torque templates for assistance, but these methods are difficult to adapt to the biomechanical differences between individuals and cannot dynamically reflect the user's real needs at different speeds, slopes, and turns. Even though torque estimation based on inverse dynamics can provide some reference, the prediction results often fluctuate greatly due to noise interference and real-time issues. Without a continuous and smooth mapping mechanism, direct output to the actuator can easily generate unstable impact torques, thereby affecting the wearer's experience and system safety.

[0005] Third, the recognition accuracy of single-modal signals in complex motion is insufficient. Attitude estimation based on inertial measurement units (IMUs) is susceptible to drift and magnetic interference; while intention recognition based on electromyography (EMG) signals is intuitive, it is limited by individual differences, noise, and electrode attachment conditions, making long-term stable application difficult; gait detection based on plantar pressure (FP) shows decreased recognition accuracy in non-flat terrain or rapid speed changes. All single modalities suffer from insufficient robustness and universality, making it difficult to support exoskeletons in achieving reliable phase estimation and pattern recognition in complex scenarios.

[0006] In summary, existing exoskeleton control methods suffer from deficiencies in gait phase estimation continuity, motion pattern recognition reliability, and the realism and stability of torque actuation. Therefore, there is an urgent need to propose an exoskeleton control method that integrates multimodal signals: replacing discrete event triggering with continuous gait phase modeling to ensure smoothness and robustness of control; simultaneously introducing realistic mechanical constraints and individualized torque estimation mechanisms, combined with a smooth mapping strategy, to achieve natural and stable human-machine collaborative control, significantly improving the adaptability and user experience of exoskeletons in complex environments. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, this invention provides an exoskeleton driving method based on real torque constraints and continuous gait phase modeling. This method solves the problems of unstable gait phase recognition, unsmooth mode switching, and lack of continuity and individualized adaptability in torque output of existing exoskeletons in complex environments.

[0008] To achieve the aforementioned objectives, the present invention employs the following technical solution: an exoskeleton driving method based on real torque constraints and continuous gait phase modeling, comprising: Multimodal signal data during the subject's gait process were acquired using a multi-sensor synchronous acquisition system, including IMU signals, EMG signals, plantar pressure signals, and optical motion capture data. The multimodal signal data is preprocessed to obtain the preprocessed signal; Continuous gait phase prediction and gait pattern recognition are performed based on the preprocessed signal; Instantaneous joint torques are estimated using a pre-trained torque estimation network based on the predicted gait phase and gait pattern. The obtained instantaneous joint torque is converted into a drive signal for the exoskeleton actuator.

[0009] The beneficial effects of this invention are as follows: 1. Improved accuracy and robustness of gait recognition. By fusing inertial measurement unit (IMU), electromyography (EMG) signals, and plantar pressure data, a continuous phase constraint method is employed to achieve accurate prediction of gait phase. The complementarity of multimodal signals effectively overcomes the problems of single-sensor susceptibility to environmental interference and insufficient accuracy, enabling the system to maintain stable recognition performance under complex motion conditions such as different speeds, inclines, declines, and turns.

[0010] 2. High-precision estimation of joint torque is achieved. This invention introduces the real torque obtained by inverse dynamics calculation as a training label, and combines it with multimodal signals to predict the instantaneous torque of the hip and knee joints. This breaks through the limitations of traditional methods that rely solely on kinematic parameters or single sensors for inference, and significantly improves the reliability and individualized adaptability of torque estimation in exoskeleton control.

[0011] 3. Ensures the smoothness and naturalness of exoskeleton control. By designing a continuous torque mapping mechanism, the predicted instantaneous joint torque is smoothly converted into exoskeleton drive signals, avoiding the jitter and delay problems caused by discrete control. This achieves a natural and stable mechanical assistance effect, improving comfort and safety during use.

[0012] 4. Enhanced system adaptability and scalability. The multimodal fusion framework of this invention is not only applicable to different groups of people and diverse motion scenarios, but also has the potential to be extended to more sensor signals, which can further expand its application scope in rehabilitation medicine, assisted walking, and human-machine collaborative robots.

[0013] In summary, this invention has achieved breakthrough improvements in gait phase prediction, torque estimation, and control smoothness, which not only enhances technical performance but also significantly improves the human-computer interaction experience. It has broad application value in various scenarios such as rehabilitation training, sports assistance, and intelligent exoskeletons. Attached Figure Description

[0014] Figure 1 A flowchart of an exoskeleton driving method based on real torque constraints and continuous gait phase modeling is provided for an embodiment. Figure 2 This is a schematic diagram of the gait phase prediction module; Figure 3 Flowchart of the method for obtaining actual joint torque; Figure 4 This is a schematic diagram of the torque estimation process. Detailed Implementation

[0015] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0016] like Figure 1 As shown, in one embodiment of the present invention, the exoskeleton driving method based on real torque constraints and continuous gait phase modeling includes the following steps: S1. Acquire multimodal signal data during the subject's gait using a multi-sensor synchronous acquisition system, including IMU signals, EMG signals, plantar pressure signals, and optical motion capture data.

[0017] This study selected 30 healthy adult subjects as research participants. All subjects were young adults with no history of lower limb injury, surgery, or gait disorder, aged between 20 and 30 years old, with heights ranging from 160 to 190 cm and weights ranging from 50 to 100 kg.

[0018] Subjects wore shoes equipped with inertial measurement units (IMUs), electromyography (EMG) sensors, and plantar pressure sensors (FPs), and reflective markers were attached to the surfaces of lower limb joints and major motor segments to simultaneously acquire human joint kinematic information via an optical motion capture system. An optical motion capture camera tracked 36 clinically standardized markers on the lower limbs at a sampling frequency of 240 Hz to obtain high-precision lower limb kinematic data. Simultaneously, subjects completed gait experiments on a three-dimensional force measurement platform to synchronously record three-dimensional ground reaction forces during gait.

[0019] Gait testing includes not only subjects walking at their own speed on level ground, but also a variety of complex movement scenarios, such as different speeds (slow walking, normal walking, fast walking), walking uphill and downhill, and turning, to ensure that the collected data is diverse and widely adaptable.

[0020] Each subject completed 10 gait tests in different scenarios, and a total of 300 gait samples were collected (30 subjects × 10 tests), which included synchronized IMU signals, EMG signals, plantar pressure signals and optical motion capture data.

[0021] S2. Preprocess the multimodal signal data to obtain the preprocessed signal.

[0022] To ensure the consistency and reliability of multimodal signals, this invention performs multi-channel preprocessing and feature extraction on the acquired data: Synchronization Alignment: Using hardware trigger signals or a unified timestamp, time alignment is achieved between the IMU, EMG, plantar pressure sensor, 3D force stage and optical motion capture system to ensure multimodal signal synchronization.

[0023] IMU signal processing: Zero bias correction is performed on accelerometer and gyroscope data, and complementary filtering or extended Kalman filtering is used to fuse and calculate the posture information of each movement segment of the lower limb; the joint angle curves of the hip and knee joints are obtained by the posture differences between adjacent segments.

[0024] EMG signal processing: The raw signal is bandpass filtered at 20–450 Hz to remove noise and drift, and then full-wave rectification and low-pass filtering at 3–6 Hz are performed to extract the muscle activation envelope; after normalization, principal component analysis (PCA) or neural network (NN) methods are used to extract multi-muscle group synergistic features.

[0025] Plantar pressure (FP) signal processing: The sensor data is calibrated, and the total pressure and the load ratio in the front-to-back and left-to-right directions are calculated; the plantar pressure center (COP) trajectory and ground contact probability are further extracted to characterize the plantar force pattern under different time states.

[0026] S3. Perform continuous gait phase prediction and gait pattern recognition based on the preprocessed signal.

[0027] like Figure 2 As shown, the specific method for continuous gait phase prediction is as follows: First, continuous phase modeling is performed.

[0028] A complete gait cycle is defined as a continuous interval [0, 2π], and an orthogonal phase modeling method based on multimodal signals is proposed. Specifically, the preprocessed IMU signal, EMG signal, and plantar pressure signal are considered as mutually independent orthogonal components, and a three-dimensional signal vector is constructed, the expression of which is:

[0029] in, Represents a three-dimensional signal vector. This represents the trajectory of the plantar pressure center or key feature components, i.e., the preprocessed plantar pressure signal; This represents the electromyographic features after envelope extraction and filtering, i.e., the preprocessed EMG signal; This represents the attitude or acceleration characteristics of the inertial measurement unit, i.e., the pre-processed IMU signal; Normalizing the three-dimensional signal vector yields a three-dimensional temporal trajectory located on a unit sphere, which forms a closed loop throughout the gait cycle. The obtained three-dimensional temporal trajectory is mapped to continuous gait phase variables, specifically as follows: Principal component decomposition is performed on the three-dimensional signal vector to extract two-dimensional sub-vectors of the three-dimensional signal vector on the first principal component and the second principal component; where the first principal component is the direction with the largest variance and the second principal component is the direction orthogonal to the first principal component and with the second largest variance. A unified gait phase is defined based on the extracted two-dimensional sub-vectors, and its expression is as follows:

[0030] in, This represents the defined uniform gait phase, i.e., the continuous gait phase; It is the arctangent function. Indicated in the first principal component The two-dimensional sub-vectors extracted from the above, Indicated in the second principal component The two-dimensional sub-vector extracted from the above.

[0031] The above steps ensure the mathematical continuity and physical multimodal consistency of gait phase. Compared with traditional phase construction methods based on only a single signal, orthogonal modeling can simultaneously integrate information from mechanics, neural control, and kinematics, better reflecting the true gait rhythm of the human body in complex environments.

[0032] Secondly, multimodal fusion gait phase estimation is performed.

[0033] The optimization objective is defined to adjust the parameters of the gait phase estimation network, and its expression is:

[0034] in, Describe the objective function. To represent the square of the Euclidean norm, This represents the weighting factor, used to balance prediction accuracy and phase monotonicity. This represents the predicted gait phase value at time t+1.

[0035] Feature extraction is performed on the preprocessed IMU signal, EMG signal, and plantar pressure signal, and then they are fused to obtain a fused feature vector, the expression of which is:

[0036] in, Represents the fused feature vector. This represents the features extracted from the preprocessed plantar pressure signal. This represents the features extracted from the preprocessed EMG signal. This represents the features extracted from the preprocessed IMU signal; The fused feature vectors are input into the gait phase estimation network after parameter adjustment for gait phase prediction. The expression is as follows:

[0037] in, for t Gait phase prediction at time t, This is the corresponding processing for the gait phase estimation network. The gait phase estimation network is a temporal convolutional network that includes temporal convolutional layers and a dynamic attention mechanism.

[0038] In gait pattern recognition, this invention fuses the aforementioned phase modeling results with multimodal temporal features, and further utilizes these multimodal features for gait pattern classification, recognizing complex scenarios such as fast walking, slow walking, uphill / downhill walking, and turning. The classification model for gait pattern recognition is as follows:

[0039] in, Represents the fused feature vector Belongs to the gait pattern category The probability, For activation function, For the weights of the classification model, This is the bias term for the classification model. This indicates the gait pattern category, including but not limited to walking at a constant speed, walking briskly, walking slowly, walking uphill, walking downhill, and turning.

[0040] S4. Based on the predicted gait phase and gait pattern, the instantaneous torque of the joint is estimated using a pre-trained torque estimation network.

[0041] This invention, based on an optical motion capture system and a ground reaction force measurement system, calculates the true biomechanical torques of the human hip, knee, and ankle joints. These torques are then used as high-precision supervised labels input to a torque estimation network, thereby achieving accurate estimation of the instantaneous torques of the hip, knee, and ankle joints during the gait cycle. Furthermore, this invention utilizes pre-predicted gait pattern categories to modulate the torque estimation results, improving the accuracy and adaptability of the estimation. The torque estimation network is an instantaneous convolutional network (TCN).

[0042] like Figure 3 As shown, the specific method for obtaining the actual joint torque is as follows: A personalized musculoskeletal model was built using OpenSim. Based on a general human template, the model was scaled according to the subject's height, weight, and lower limb bone segment length.

[0043] in, Scaling factor The length of the subject's bone segment, This is the length corresponding to the template model.

[0044] The anatomical center positions of the hip, knee, and ankle joints in the individualized musculoskeletal model are calibrated using the marker point trajectory. The calibration function is as follows:

[0045] in, For calibration function, The coordinates of the marker points collected in the experiment. The coordinates predicted by the model. Indicates joint parameters, This represents the total number of markers used for calibration.

[0046] The joint angle curve is calculated using the trajectory of marked points, and its expression is as follows:

[0047] in, This represents the optimized joint angle vector. Weights for the marker points; This represents the coordinates of the i-th marker point collected at time t. This indicates that, given a joint angle vector Below, the coordinates of the i-th marker point predicted by the model. This represents the candidate joint angle vector to be optimized; The complete kinematic time series characteristics are obtained by calculating angular velocity and angular acceleration using numerical differentiation, and their expression is as follows:

[0048] in, Indicates joint angular velocity, This represents the joint angular acceleration.

[0049] Combining the obtained kinematic time-series characteristics with ground reaction force data, and substituting them into the inverse dynamic equation, the expression is as follows:

[0050] in, For the quality matrix, For Coriolis force and centrifugal force terms, For gravity, This is the net torque of the joint. This is the transpose of the Jacobian matrix. The ground reaction force; The net torque of the joints is obtained by solving the inverse dynamics equation, thereby obtaining the net torque of the hip, knee and ankle joints in the sagittal plane, which is the true value of the biological joint torque. To ensure comparability between different individuals, this invention normalizes the calculated torque, and its expression is as follows:

[0051] in, This represents the normalized true value of the biological joint torque. The true values ​​of biological joint torques before normalization. represents weight, It is the acceleration due to gravity. This is a standardized reference lower limb length.

[0052] Standardized torque labels were obtained, and then IMU, plantar pressure, and electromyography signals were aligned with the obtained torque labels along the temporal dimension, strictly matched on the same timeline, ultimately resulting in a standard multimodal lower limb motion dataset with torque labels. These torque labels not only reflect the actual biomechanical requirements of the human body under complex scenarios such as different speeds, slopes, and turns, but also provide reliable monitoring signals for exoskeleton systems.

[0053] The loss function used in supervised training of the torque estimation network is:

[0054] in, Indicates the loss value. and All are weighting factors. This is a regularization term used to prevent overfitting; As a continuity constraint, it is used to improve the smoothness and stability of the predicted torque in the time dimension.

[0055] like Figure 4 As shown, the trained torque estimation network is used to predict the instantaneous torque of the joint. The specific method is as follows: Construct the input signal, which includes the plantar pressure feature vector. Surface electromyography feature vectors of key muscle groups and the feature vector of the lower limb inertial measurement unit. The time dimension is used to combine these features into a fused feature vector, which is expressed as follows:

[0056] in, Indicates the input signal; The predicted continuous gait phases obtained from the above steps are used to provide a unified timing reference.

[0057] The input signal is combined with the gait pattern category and used as input to a pre-trained torque estimation network to obtain the predicted instantaneous joint torque, which is expressed as follows:

[0058] in, This represents the predicted instantaneous torque value of the joint. The mapping function represents the moment estimation network. This represents the encoding vector corresponding to the gait pattern category.

[0059] , representing the predicted torques of the hip, knee, and ankle joints, respectively, in Newton-meters (N·m). By introducing... The network can dynamically adjust torque estimation based on gait pattern categories. For example, it can appropriately increase the predicted torque of hip and knee extensors in uphill mode and decrease the predicted torque of ankle dorsiflexion in downhill mode, thus better reflecting actual movement patterns.

[0060] Through the above steps, this invention not only enables continuous gait phase and pattern recognition under multimodal input, but also achieves high-precision estimation of instantaneous joint torque by combining gait pattern categories within the same framework. This method ensures the accuracy of torque estimation while improving the system's adaptability to different motion environments through pattern-based adjustments, thereby providing precise reference signals and reliable control basis for the torque drive of the exoskeleton system.

[0061] S5. Convert the obtained instantaneous joint torque into a drive signal for the exoskeleton actuator.

[0062] To accommodate different body weights, joint biomechanical characteristics, and usage scenarios (such as rehabilitation training and exercise assistance), the predicted instantaneous torque value of the joint is scaled, and its expression is as follows: ,

[0063] in, This is the predicted instantaneous torque value of the joint after amplitude scaling. This is a scaling factor used to adjust the intensity of the assistance. A smaller scaling factor can be selected in the early stages of rehabilitation. To avoid over-intervention; in motion-assisted scenarios, the intensity can be appropriately increased according to user needs. This enhances output. The scaling factor can be set manually, calculated based on weight-normalized rules, or automatically adjusted using an adaptive algorithm.

[0064] Due to varying degrees of delay in multimodal signal acquisition, TCN network inference, and exoskeleton actuator response, without compensation, the exoskeleton assistance may become out of sync with the actual human movement, resulting in abruptness or awkwardness. To address this issue, a time compensation term is introduced to compensate for the delay in the amplitude-scaled predicted instantaneous joint torque. Its expression is as follows:

[0065] in, This is the predicted instantaneous torque value of the joint after time delay compensation. This is a time compensation item. The value of can be obtained through experimental calibration, and is usually within the range of 10–50 ms. Other embodiments of the present invention may also employ adaptive time delay estimation algorithms (such as cross-correlation alignment and dynamic time warping) to achieve individualized time delay compensation, further ensuring that the exoskeleton output is strictly in phase with the human body's force exertion process.

[0066] To avoid torque spikes caused by noise, EMG transients, or algorithm prediction errors, and to maintain smoothness and continuity while eliminating high-frequency jitter, this embodiment uses a second-order Butterworth low-pass filter to smooth the compensated torque signal. The expression for this filter is:

[0067] in, The predicted instantaneous torque of the joint after smoothing. Let be the impulse response function of the low-pass filter, with the symbol . This represents the convolution operation; In addition to the basic processing methods mentioned above, the predicted instantaneous torque values ​​of the joints after smoothing can be optimized to obtain the drive signals of the exoskeleton actuators.

[0068] The optimization of the smoothed joint instantaneous torque prediction value includes: Set upper and lower threshold values ​​for the predicted instantaneous torque of the joint to ensure that the drive signal is always within a safe range and avoid excessive torque causing secondary damage to the human body; According to gait phase Dynamically adjust the scaling factor The value of is used to implement a phase-dependent personalized assist strategy, and its expression is:

[0069] in, A mapping function, used to map... From the interval Mapping to interval , This represents the radian value of a 180-degree angle within the unit circle. This represents the minimum value of the scaling factor. This represents the maximum value of the scaling factor; thus achieving a larger assist output in the support phase and a smaller assist output in the swing phase.

[0070] A rule-based control strategy is used to generate a compensation torque, and the smoothed instantaneous torque prediction of the joint is fused with the generated compensation torque.

[0071] Through the aforementioned amplitude scaling, delay compensation, low-pass filtering, and optimized continuous torque mapping mechanism, this invention effectively achieves a natural conversion from predicted biological joint torque to exoskeleton driving torque. This ensures that the exoskeleton's auxiliary torque remains strictly synchronized with the actual force exertion process of the human body in the time domain, and achieves reasonable constraints in the amplitude and frequency domains, thereby improving the naturalness and safety of human-machine collaboration.

[0072] In summary, this invention has achieved breakthrough improvements in gait phase prediction, torque estimation, and control smoothness, which not only enhances technical performance but also significantly improves the human-computer interaction experience. It has broad application value in various scenarios such as rehabilitation training, sports assistance, and intelligent exoskeletons.

Claims

1. An exoskeleton driving method based on real torque constraint and continuous gait phase modeling, characterized in that, The application relates to a gait analysis method and system. The method comprises the following steps: acquiring multi-modal signal data of a subject during gait by using a multi-sensor synchronous acquisition system, including IMU signals, EMG signals, plantar pressure signals and optical motion capture data; preprocessing the multi-modal signal data to obtain preprocessed signals; predicting continuous gait phases and recognizing gait patterns according to the preprocessed signals; estimating joint instantaneous torque based on the predicted gait phases and gait patterns by using a pre-trained torque estimation network; 2. The method of claim 1, wherein, converting the obtained joint instantaneous torque into a driving signal of an exoskeleton actuator. The specific method for predicting continuous gait phases according to the preprocessed signals is as follows: modeling continuous phases according to the preprocessed IMU signals, EMG signals and plantar pressure signals: wherein, represents a three-dimensional signal vector, represents a plantar pressure center trajectory or key feature component, represents an envelope-extracted and filtered electromyography feature, represents an inertial measurement unit's attitude or acceleration feature; regarding the preprocessed IMU signals, EMG signals and plantar pressure signals as mutually independent orthogonal components, and constructing a three-dimensional signal vector, the expression of which is as follows: normalizing the three-dimensional signal vector to obtain a three-dimensional time sequence trajectory on a unit sphere, which forms a closed loop in the whole gait cycle; mapping the obtained three-dimensional time sequence trajectory into a continuous gait phase variable, the specific method being as follows: performing principal component decomposition on the three-dimensional signal vector to extract a two-dimensional sub-vector of the three-dimensional signal vector on the first principal component and the second principal component; wherein the first principal component is the direction with the largest variance, and the second principal component is the direction orthogonal to the first principal component and with the second largest variance; wherein denotes a defined uniform gait phase, i.e. a continuous gait phase; is an arc tangent function, denotes a two-dimensional sub-vector extracted on the first principal component denotes a two-dimensional sub-vector extracted on the second principal component denotes a two-dimensional sub-vector extracted on the first principal component denotes a two-dimensional sub-vector extracted on the second principal component defining a unified gait phase according to the extracted two-dimensional sub-vector, the expression of which is as follows: wherein, represents a target function, represents a square of the Euclidean norm, represents a weight factor, is a gait phase prediction value at time t+1, is t a gait phase prediction value at time t. defining an optimization objective to adjust the parameters of the gait phase estimation network, the expression of which is as follows: wherein, represents a fusion feature vector, represents a feature extracted from a pre-processed plantar pressure signal, represents a feature extracted from a pre-processed EMG signal, represents a feature extracted from a pre-processed IMU signal; performing feature extraction on the preprocessed IMU signals, EMG signals and plantar pressure signals, and then fusing the features to obtain a fused feature vector, the expression of which is as follows: wherein, is a mapping function for the gait phase estimation network.

3. The method of claim 2, wherein, inputting the fused feature vector into the gait phase estimation network with adjusted parameters to predict the gait phase, the expression of which is as follows: wherein, represents a gait pattern class, represents a fused feature vector belongs to a gait pattern class is an activation function, is a weight of the classification model, is a bias term of the classification model.​ 4. The method of claim 2, wherein, the classification model for recognizing gait patterns according to the preprocessed signals is as follows: supervised training the torque estimation network by using real joint torque as a supervision label; the method for obtaining the real joint torque is as follows: scaling an individualized musculoskeletal model according to the height, weight and lower limb bone segment length of the subject based on a general human template; wherein, is a calibration function, is a coordinate of a marker point collected in an experiment, is a coordinate predicted by a model, denotes a joint parameter, is a total number of marker points used for calibration; calibrating the anatomical center positions of the hip, knee and ankle joints in the individualized musculoskeletal model by using marker point trajectories, the calibration function being as follows: wherein, denotes the optimized joint angle vector, is a marker point weight; denotes the coordinate of the i-th marker point collected at time t, denotes the given joint angle vector the coordinate of the i-th marker point predicted by the model, denotes the candidate joint angle vector to be optimized; calculating a joint angle curve by using the marker point trajectories, the expression of which is as follows: wherein, denotes the joint angular velocity, denotes the joint angular acceleration; calculating angular velocity and angular acceleration by numerical differentiation to obtain complete kinematic time sequence characteristics, the expression of which is as follows: wherein is a mass matrix, is a Coriolis and centrifugal force term, is a gravitational force term, is a joint net force moment, is the transpose of the Jacobian matrix, is the ground reaction force; combining the obtained kinematic time sequence characteristics and ground reaction force data, and substituting them into an inverse dynamics equation, the expression of which is as follows: solving the joint net torque according to the inverse dynamics equation to obtain the net torque of the hip, knee and ankle joints in the sagittal plane, i.e. the real value of the biological joint torque; wherein, is the normalized biological joint torque true value, is the non-normalized biological joint torque true value, denotes the body weight, is the gravitational acceleration, is the normalized reference lower limb length.

5. The method of claim 4, wherein, normalizing the obtained real value of the biological joint torque, the expression of which is as follows: the specific method for estimating joint instantaneous torque is as follows: wherein represents an input signal; constructing an input signal, the expression of which is as follows: combining the input signal and the gait pattern category as the input of the pre-trained torque estimation network to obtain a joint instantaneous torque prediction value, the expression of which is as follows: wherein, represents a joint instantaneous torque prediction value, represents a corresponding processing of the torque estimation network, represents an encoding vector corresponding to the gait pattern class.

6. The method of claim 5, wherein, The loss function used in the supervised training of the torque estimation network is: wherein, represents a loss value, and are weight factors, is a regularization term for preventing overfitting; is a continuity constraint for improving the smoothness and stability of the predicted moment in the time dimension.

7. The method of claim 6, wherein, The specific method for converting the obtained joint instantaneous torque into a driving signal of the exoskeleton actuator is: The amplitude scaling of the joint instantaneous torque prediction value is expressed as: ， wherein, is the scaled joint instantaneous torque prediction value, is the scaling factor; The time compensation term is introduced to compensate for the delay of the amplitude-scaled joint instantaneous torque prediction value, and the expression is: wherein, is the time-compensated joint instantaneous torque prediction value, is the time compensation term; The smoothed joint instantaneous torque prediction value is expressed as: wherein is the smoothed joint instantaneous torque prediction value, is an impulse response function of a low-pass filter, and the symbol denotes a convolution operation; The smoothed joint instantaneous torque prediction value is optimized to obtain the driving signal of the exoskeleton actuator.

8. The method of claim 7, wherein, The optimization of the smoothed joint instantaneous torque prediction value includes: Setting upper and lower threshold values for the joint instantaneous torque prediction value to ensure that the driving signal is always within a safe range; According to gait phase Dynamic adjustment of the scale factor The value of the scale factor is dynamically adjusted according to the gait phase, and the expression of the phase-dependent individual assistance strategy is: wherein is a mapping function for mapping from the interval to the interval , denotes the radian value of a 180-degree angle in the unit circle, denotes the minimum value of the scale factor, denotes the maximum value of the scale factor; Using a rule-based control strategy to generate a compensation torque, and fusing the smoothed joint instantaneous torque prediction value with the generated compensation torque.

9. The method of claim 1, wherein, The pre-processing of the multi-modal signal data is specifically: Using hardware trigger signals or uniform timestamps to align the time between IMU signals, EMG signals, plantar pressure signals, and optical motion capture data; For IMU signals, zero offset correction is performed on accelerometer and gyroscope data, and attitude information of each motion segment of the lower limb is calculated by complementary filtering or extended Kalman filtering fusion; the joint angle curves of the hip and knee joints are obtained by the attitude difference between adjacent segments; For EMG signals, band-pass filtering is used to remove noise and drift, and full-wave rectification and low-pass filtering are used to extract muscle activation envelopes; the muscle activation envelopes are normalized, and multi-muscle group coordination features are extracted using principal component analysis or neural network methods; For plantar pressure signals, sensor data is calibrated, total pressure and front-back and left-right load ratios are calculated; plantar pressure center trajectory and ground contact probability are extracted to describe the plantar force mode under different gait conditions.

Images