Lower limb motion intention prediction method and system based on myoelectric-inertial signals

Abstract

This invention belongs to the field of signal processing technology, specifically relating to a method and system for predicting lower limb movement intention based on electromyography (EMG) and inertial signals. The method includes: acquiring surface EMG signals from multiple muscle groups in the lower limb and posture angle data from an inertial measurement unit, and preprocessing the surface EMG signals by filtering; extracting time-domain and / or frequency-domain features from the preprocessed signals to construct an EMG feature sequence; inputting the EMG features into a lightweight multi-scale convolutional attention network to identify the current lower limb movement pattern; simultaneously inputting the EMG features into a channel attention-enhanced long short-term memory network to predict the target angles of the hip, knee, and ankle joints; based on the posture angle data, modeling and calculating the actual angles of each joint in the lower limb using a forward kinematics D-H coordinate system, and comparing them with the target angles to obtain the angle deviation; finally, based on the angle deviation and the identified movement pattern, generating control commands to adjust the output torque of the exoskeleton joint actuators to achieve compliant following assistance.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a method and system for predicting lower limb movement intentions based on electromyography-inertial signals. Background Technology

[0002] With the accelerating aging process in my country, the problem of motor dysfunction among the elderly is becoming increasingly prominent. According to data released by the National Committee on Aging in 2023, there are approximately 26.28 million elderly people in my country currently suffering from mild motor dysfunction, and this number is continuing to grow at a rate of 3%-5% annually, projected to exceed 35 million by 2030. This serious situation makes the development of assistive devices for the elderly with mild motor dysfunction a crucial issue in the field of health technology.

[0003] The decline in lower limb motor function is particularly prominent in the physiological decline of the elderly. Medical research shows that approximately 42% of people over 65 years of age experience varying degrees of quadriceps muscle weakness, 31% suffer from proprioceptive dysfunction, and 28% experience vestibular system degeneration. These physiological changes pose significant challenges for the elderly when performing daily posture transitions (such as standing up and sitting down). Clinical statistics show that more than 30 million falls occur annually among people over 65 years of age in my country, with 72.6% occurring during the transition between sitting and standing, resulting in serious consequences such as hip fractures and increasing medical expenditures by nearly 20 billion yuan. Traditional lower limb exoskeleton care robots suffer from problems such as excessive motion load, fixed motion trajectory, and lack of human-machine posture coordination, necessitating a posture transition assistive robot for the elderly that can provide real-time feedback and intelligent adaptation.

[0004] Existing lower limb exoskeleton assistive robots for the elderly mostly use traditional CNN or LSTM models for electromyographic intention recognition, which have problems such as poor noise robustness and insufficient extraction of multi-scale spatiotemporal features, making it difficult to adapt to the weak electromyographic signals of the elderly. The control link is mostly open-loop control with preset trajectory, without collecting the actual joint movement state, lacking real-time feedback and deviation correction, and unable to dynamically adjust the assist torque and timing. Poor human-machine coordination and insufficient following compliance result in poor assistive effect and cannot meet the assistance needs of the elderly for safety and comfort.

[0005] Therefore, it is very important to design a method and system for predicting lower limb movement intentions based on electromyography-inertial signals that can perform well in motion intention recognition and joint coordination control, and effectively improve the following and comfort of elderly exoskeletons. Summary of the Invention

[0006] This invention aims to overcome the shortcomings of existing lower limb exoskeleton assistive robots for the elderly, which often use traditional CNN or LSTM models for electromyographic intention recognition. These models are difficult to adapt to the weak electromyographic signals of the elderly and suffer from poor noise robustness and insufficient extraction of multi-scale spatiotemporal features. The invention provides a method and system for predicting lower limb movement intention based on electromyographic-inertial signals that can perform well in motion intention recognition and joint coordination control, and effectively improve the following performance and comfort of the elderly assistive exoskeleton.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0008] A method for predicting lower limb movement intention based on electromyography-inertial signals includes the following steps:

[0009] S1, Signal Acquisition and Preprocessing: Acquire surface electromyography (EMG) signals from multiple muscle groups in the lower limbs, as well as attitude angle data from the lower limb inertial measurement unit, and perform filtering preprocessing on the surface EMG signals;

[0010] S2, Electromyography Feature Extraction: Extract time-domain and / or frequency-domain features from the preprocessed surface electromyography signal to construct an electromyography feature sequence;

[0011] S3, Movement Intent Recognition: The extracted electromyographic features are input into a lightweight multi-scale convolutional attention network to identify the current lower limb movement pattern;

[0012] S4, Joint Angle Prediction: The extracted electromyographic features are input into a channel attention-enhanced long short-term memory network to predict the target angles of the hip, knee, and ankle joints.

[0013] S5, Position Calculation and Cooperative Assist Control: Based on the posture angle data, the actual angles of each joint of the lower limb are calculated by modeling in the forward kinematics DH coordinate system and compared with the target angle to obtain the angle deviation; based on the angle deviation and the identified lower limb movement pattern, control commands are generated to adjust the output torque of the actuator of the exoskeleton joint.

[0014] Preferably, in step S1, the multiple muscle groups of the lower limbs include:

[0015] The gluteus maximus, gluteus medius, gluteus minimus, and iliopsoas muscles in the hip region;

[0016] The quadriceps femoris, hamstrings, lower gluteus maximus, and gracilis muscles of the thigh;

[0017] The gastrocnemius muscle (medial and lateral heads), soleus muscle, tibialis anterior muscle, and peroneus longus muscle of the lower leg and foot;

[0018] The electrodes for the surface electromyography (EMG) signal are arranged along the direction of the muscle fibers of each muscle group and cover the muscle belly region of the corresponding muscle group.

[0019] Preferably, in step S1, the filtering preprocessing employs a combined noise reduction strategy, as follows:

[0020] The Butterworth fourth-order bandpass filter with a passband range of 20Hz-450Hz was used, as well as multi-stage Butterworth notch filters with center frequencies of 50Hz, 100Hz and 150Hz.

[0021] The amplitude and frequency characteristics of the Butterworth filter satisfy the following relationship: ;

[0022] In the formula, n is the filter order. The cutoff frequency, For frequency; This is the frequency response function of the Butterworth filter.

[0023] Preferably, in step S2, the time-domain features include at least one of the following: mean absolute value, root mean square, variance, wavelength, zero-crossing rate, autoregressive coefficient, and integrated electromyographic value; the frequency-domain features include at least one of the following: mean frequency, median frequency, and power spectral density.

[0024] The formula for calculating the integrated electromyographic value is as follows: ;

[0025] In the formula, It is the electromyographic signal voltage value at time t; , These are the start and end times of the analysis time window, respectively. It is the absolute integral value of the electromyographic signal within the time window, used to reflect the total amount of muscle activation;

[0026] The median frequency (MDF) satisfies: ;

[0027] In the formula, It is the power spectral density value at frequency f; It is the maximum frequency of the analysis band.

[0028] Preferably, in step S3, the lightweight multi-scale convolutional attention network includes adaptive one-dimensional depthwise separable convolution and a dynamic and efficient channel attention mechanism.

[0029] The adaptive one-dimensional depthwise separable convolution is based on the average frequency of the input electromyographic features. The adaptive selection of convolutional kernel size, and the corresponding processing steps are as follows: ; ;

[0030] The original electromyographic feature matrix is ​​input into an adaptive one-dimensional depthwise separable convolution. L is the signal time step, and C is the number of electrode channels. Represents the real number field; This indicates that a spatial convolution operation is performed independently for each input channel; The output features are obtained by adaptive one-dimensional depthwise separable convolution + pointwise convolution; PointwiseConv = pointwise convolution; The characteristic frequency of the electromyographic signal at time t;

[0031] The dynamic and efficient channel attention mechanism integrates a signal-to-noise ratio feedback adjustment mechanism to dynamically generate channel weights W. The corresponding processing procedure is as follows: ; ; ;

[0032] in, The sigmoid activation function is used, SNR(X) is the signal-to-noise ratio of each channel, and SNRmax=40dB is the preset normalization threshold; X is the original feature matrix, Y is the enhanced feature matrix; γ=2, b=1; This indicates element-wise multiplication; GlobalAvgPool represents a one-dimensional sliding convolution on the temporal dimension to extract local features and their association with channels, which are then used to generate channel attention weights. represents the size of the one-dimensional convolution kernel; odd indicates taking the nearest odd number from the calculated result.

[0033] Preferably, in step S4, the channel attention-enhanced long short-term memory network includes a feature extraction stage, a feature refinement stage, and a prediction output stage;

[0034] The feature extraction stage extracts the temporal dependencies of electromyographic features through a one-dimensional convolutional neural network embedding layer and a long short-term memory network layer.

[0035] The feature refinement stage includes a fine-grained channel attention module and a frequency-domain enhanced channel attention module. The fine-grained channel attention module generates channel weights through depthwise convolution, pointwise convolution, and global pooling. The frequency-domain enhanced channel attention module transforms the time-domain signal to the frequency domain through a one-dimensional discrete cosine transform to extract spectral features. The transformation formula is as follows: ;

[0036] in, The length of a single-channel signal is given by l, where l is the frequency component index, i.e., from 0 to l. -1; These are time-domain signal sampling points; is the frequency domain component obtained after the one-dimensional discrete cosine transform; i is the sampling point index of the time domain signal;

[0037] The prediction output stage aggregates temporal context information through a historical feature fusion module and uses a temporal pattern attention network to capture the impact of historical moments on the current angle prediction.

[0038] Preferably, in step S5, the actual angles of each joint of the lower limb are calculated by modeling using the positive kinematics DH coordinate system, which includes the following process:

[0039] Based on the linkage model of the lower limb and the DH coordinate system established at each joint, the spatial position of the lower limb distal end in the base coordinate system is calculated through homogeneous coordinate transformation; for the metatarsal node D, the corresponding coordinates in the reference coordinate system {O} are... The solution is obtained using the following formula: ;

[0040] in, Let D be the coordinates of the metatarsal node D in coordinate system {D}; Let be the homogeneous transformation matrix from coordinate system {C} to coordinate system {D}; Let {B} be the homogeneous transformation matrix from coordinate system {B} to coordinate system {C}. Let be the homogeneous transformation matrix from coordinate system {A} to coordinate system {B}; Let be the homogeneous transformation matrix from coordinate system {O} to coordinate system {A}.

[0041] Preferably, the training process of the lightweight multi-scale convolutional attention network adopts a strategy of independent partitioning based on the subjects, and the subjects included in the training set, validation set, and test set do not overlap, and are divided into training set, validation set, and test set according to a preset ratio; the network training is optimized using the cross-entropy loss function, the expression of which is: ;

[0042] Where N is the number of training samples per batch, and M is the number of motion pattern categories. Let i be the true label of the i-th sample in the c-th motion pattern. To predict the probability that the i-th sample belongs to the c-th motion pattern.

[0043] Preferably, in step S5, the lower limb movement pattern includes one or more of standing up, sitting down, walking on flat ground, and turning left; the actuator is a pneumatic muscle actuator, and the relationship between the corresponding driving force F and the internal working pressure p satisfies: ;

[0044] in, To simplify the coefficients, D0 is the original diameter of the muscle. The initial angle between the fiber layer and the axial direction;

[0045] The relationship between joint driving torque, muscle output force, equivalent force arm, and joint rotation angle satisfies: ;

[0046] in, For joint driving torque, This refers to the amount of muscle contraction. For muscle force driving characteristic function, It is an equivalent arm.

[0047] The present invention also provides a lower limb movement intention prediction system based on electromyography-inertial signals, including:

[0048] The surface electromyography signal acquisition module includes multiple electrodes for attaching to multiple muscle groups in the lower limbs to acquire surface electromyography signals.

[0049] The inertial measurement module includes multiple inertial measurement units located in each segment of the lower limbs, used to collect attitude angle data;

[0050] Processor, used to perform the following operations:

[0051] The surface electromyography signal is filtered and preprocessed, and time-domain and / or frequency-domain features are extracted.

[0052] The extracted electromyographic features are input into a lightweight multi-scale convolutional attention network to identify the current lower limb movement pattern;

[0053] The extracted electromyographic features are input into a channel attention-enhanced long short-term memory network to predict the target angles of the hip, knee, and ankle joints.

[0054] Based on the posture angle data, the actual angles of each joint of the lower limb are calculated by modeling in the forward kinematics DH coordinate system and compared with the target angles to obtain the angle deviation;

[0055] Based on the angular deviation and the identified lower limb movement patterns, control commands are generated.

[0056] The actuator module includes multiple actuators located at the joints of the lower limb exoskeleton, which are used to receive control commands and adjust the output torque to achieve follow-alert.

[0057] Compared with the prior art, the beneficial effects of this invention are:

[0058] (1) This invention analyzes surface electromyography signals using a lightweight multi-scale convolutional attention network (LMSACNN) and predicts target angles of the hip, knee, and ankle joints using a channel attention-enhanced long short-term memory network (CAELSTM). Simultaneously, it collects actual joint angles using a lower limb inertial measurement unit, completes position calculation by modeling based on the forward kinematics DH coordinate system, and compares the calculated angles with the predicted target angles in real time. The exoskeleton control system dynamically adjusts the output torque of the pneumatic muscle actuators based on the deviation data to achieve human-machine synchronous and compliant assistance. Experimental verification shows that this invention performs excellently in motion intention recognition and joint coordination control, effectively improving the responsiveness and comfort of the elderly exoskeleton and providing reliable technical support for the elderly to live independently. Attached Figure Description

[0059] Figure 1 This is a comparative schematic diagram of the time-domain waveforms of sEMG signals (surface electromyography signals) before and after denoising when the human body is standing in the present invention; Figure 2 This is a comparative schematic diagram of the time-domain waveforms of sEMG signals (surface electromyography signals) before and after noise reduction when a human walks in a straight line, as described in this invention. Figure 3 This is a comparative schematic diagram of the time-domain waveforms of sEMG signals (surface electromyography signals) before and after noise reduction in the case of a human squatting and standing up in this invention; Figure 4 This is a comparative schematic diagram of the time-domain waveforms of sEMG signals (surface electromyography signals) before and after noise reduction when a human is sitting up in this invention. Figure 5 A bar chart illustrating the recognition accuracy of different subjects on the ENABL3S dataset in this invention; Figure 6 This is a schematic diagram illustrating the relationship between the carrier coordinate system and Euler angles in this invention; Figure 7 This is a schematic diagram of a simplified lower limb linkage model in this invention; Figure 8 This is a schematic diagram comparing the calculated trajectory and the preset trajectory in the YZ plane in this invention; Figure 9 This is a schematic diagram of an overall structure of the lower limb exoskeleton assistive robot of the present invention; Figure 10 This is a side view of the lower limb exoskeleton assistive robot of the present invention; Figure 11 This is a block diagram illustrating the control system principle of the lower limb exoskeleton assistive robot of the present invention. Figure 12 This is a flowchart of a calibration test for the pneumatic muscle actuator in this invention; Figure 13This is a schematic diagram of a curve fitting between knee joint rotation angle and rope tension under different working pressures in this invention; In the diagram: 1. Hip joint sensor; 2. Lumbar fixation device; 3. Hip joint assist component; 4. Thigh connection device; 5. Knee joint assist component; 6. Knee joint sensor; 7. Knee joint fixation device; 8. Ankle joint assist component; 9. Ankle joint fixation device; 10. Ankle joint sensor; 11. Foot support plate; 12. Detailed Implementation

[0060] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0061] This invention provides a method for predicting lower limb movement intention based on electromyography-inertial signals, comprising the following steps:

[0062] S1, Signal Acquisition and Preprocessing: Acquire surface electromyography (EMG) signals from multiple muscle groups in the lower limbs, as well as attitude angle data from the lower limb inertial measurement unit, and perform filtering preprocessing on the surface EMG signals;

[0063] S2, Electromyography Feature Extraction: Extract time-domain and / or frequency-domain features from the preprocessed surface electromyography signal to construct an electromyography feature sequence;

[0064] S3, Movement Intent Recognition: The extracted electromyographic features are input into a lightweight multi-scale convolutional attention network to identify the current lower limb movement pattern;

[0065] S4, Joint Angle Prediction: The extracted electromyographic features are input into a channel attention-enhanced long short-term memory network to predict the target angles of the hip, knee, and ankle joints.

[0066] S5, Position Calculation and Cooperative Assist Control: Based on the posture angle data, the actual angles of each joint of the lower limb are calculated by modeling in the forward kinematics DH coordinate system and compared with the target angle to obtain the angle deviation; based on the angle deviation and the identified lower limb movement pattern, control commands are generated to adjust the output torque of the actuator of the exoskeleton joint.

[0067] Specifically, step S1 includes the following process:

[0068] Electromyography signal acquisition - electrode placement:

[0069] Electrode placement for hip muscles:

[0070] The hip muscles are the starting point for power generation and the core of postural stability in lower limb movement. They dominate basic movements such as hip extension, flexion, and abduction, providing initial power for all lower limb movements, including walking, sitting up, and squatting. They also play a crucial role in pelvic stability, preventing the body's center of gravity from shifting during movement. The gluteus maximus is the core power muscle for hip extension, providing the main thrust during the push-off phase of sitting up, walking, and squatting, making it the core muscle group for high-intensity hip extension. The iliopsoas is the main hip flexor, controlling the power output for leg lifts during the swing phase of walking and leg lifts when climbing stairs, coordinating the movement between the lower limbs and trunk. The gluteus medius / gluteus minimus, as the core muscles for hip abduction, also maintain pelvic level during the support phase of walking, preventing the body from tilting to the non-supporting side, which is crucial for maintaining balance, especially important for fall prevention in the elderly. The coordinated efforts of all hip muscle groups enable flexible hip joint movement and stable pelvic support, making it the power hub for overall lower limb movement.

[0071] The electrode placement for the gluteus maximus should cover the main muscle belly region responsible for hip extension. The main electrode array is applied along an arc-shaped line on the body surface from the lateral border of the sacrum (the upper origin of the gluteus maximus) to the anterior border of the greater trochanter of the femur (near the insertion point), with a center-to-center spacing of approximately 2 cm, to capture the spatiotemporal characteristics of electromyographic signal transmission in the longitudinal direction of the muscle. To address the forward sliding of the muscle belly during large hip extension movements, an auxiliary electrode is placed approximately 2 cm proximal to the anterior border of the greater trochanter and offset forward by 1 cm. This point maintains good contact with the active muscle fibers even during hip hyperextension, ensuring signal continuity at the end of the movement.

[0072] Due to their deep location and similar functions, the gluteus medius and gluteus minimus muscles were used for coordinated data acquisition using a high-density surface electrode array. The array began approximately 5 cm below the highest point of the iliac crest (approximately at the level of the L4-L5 spinous processes), and was obliquely positioned anteroinferiorly along the muscle fiber direction (forming an angle of approximately 15° with the anterior vertical line), terminating approximately 2 cm above the tip of the greater trochanter. The array center was precisely located at the junction of the middle and posterior thirds of the line connecting the iliac crest and the tip of the greater trochanter; this area is the surface projection center of the posterior bundle of the gluteus medius (the main abduction force component). The oblique arrangement of the array, consistent with the direction of the deep muscle fibers, effectively enhanced the directional pickup of weak deep electrical signals generated during abduction movements.

[0073] Electrode placement for the iliopsoas muscle requires close proximity to the muscle belly while minimizing interference from the abdominal wall muscles. The specific location is as follows: First, locate the midpoint of the inguinal ligament (where the femoral artery pulsation can be felt), and use this as a reference point to move vertically upwards by 2 cm. Then, along the surface line connecting the anterior superior iliac spine and the pubic tubercle, move horizontally approximately 1.5 cm laterally (away from the midline of the abdomen) from this reference point. The final electrode center is located at this intersection. Below this point lies the relatively superficial anatomical window through which the iliopsoas tendon passes, successfully avoiding the dense intersection of the rectus abdominis lateral border and the internal oblique muscles. Combined with dynamic impedance matching technology, this significantly improves the purity of the hip flexion intention signal.

[0074] Electrode placement in the thigh muscles:

[0075] The thigh muscles connect the hip and knee joints and are the core force generators and coordinators of knee-hip joint movements. They dominate the extension and flexion of the knee joint, assist the extension, flexion, and adduction of the hip joint, and are also the core muscle group that maintains standing stability, directly determining the power and rhythm of lower limb movements. The quadriceps femoris is the sole dominant muscle for knee extension and a core stabilizing muscle for maintaining a standing posture. Its rectus femoris also functions as a hip flexor, enabling coordinated hip-knee movement. The vastus medialis muscle corrects patellar position at the end of knee extension, ensuring smooth movement and is central to knee extension during walking and squatting. The hamstrings are the primary muscles for knee flexion and also assist in hip extension, controlling thigh swing speed during walking and cushioning knee pressure during the squatting and landing phases, thus regulating the rhythm of flexion and extension movements. The lower gluteus maximus provides supplementary power for powerful hip extension, enhancing power output during the push-off phase of walking and the moment of force when standing up from a seated position. The gracilis, as a representative hip adduction muscle, maintains lower limb adduction stability during gait support, preventing excessive leg abduction. Through multi-dimensional force exertion and coordination, the thigh muscles achieve coordinated hip-knee movement, forming the core of power for lower limb movement.

[0076] The quadriceps femoris is the primary muscle group for knee extension and maintaining standing stability, making electrode placement crucial. The main electrode is precisely located at the midpoint of the line connecting the anterior superior iliac spine and the superior border of the patella, covering the core muscle belly of the rectus femoris and vastus intermedius. A lateral electrode is placed approximately 4 cm lateral to the quadriceps femoris, at the midpoint of the line connecting the greater trochanter and the lateral border of the patella, to monitor the synergistic activation of the vastus lateralis. A medial electrode is placed approximately 4 cm medially to the quadriceps femoris, at the upper third of the line connecting the medial femoral condyle and the superior medial border of the patella, to acquire key activity signals of the vastus medialis, especially at the end of knee extension. This three-point arrangement constitutes the most direct and stable signal source for knee extension intent.

[0077] The hamstring muscles play a dominant role in controlling knee flexion, hip extension, and the swing phase of gait. The medial sampling point is located approximately 8 cm vertically upwards from the midpoint of the posteromedial popliteal crease of the knee joint, and approximately 2 cm medial to the posterior midline of the thigh, at the common belly of the semitendinosus and semimembranosus muscles. The lateral sampling point is located approximately 7 cm vertically upwards from the midpoint of the posterolateral popliteal crease of the knee joint, and approximately 2.5 cm lateral to the posterior midline of the thigh, at the bulge of the long head belly of the biceps femoris. These two points together determine the knee flexion torque and the intention of the thigh swing during gait.

[0078] The lower gluteus maximus is the primary power source for hip extension initiation (such as standing up from a seated position) and push-off during walking. Its signal near its femoral attachment point is particularly crucial; therefore, in addition to the hip placement, an auxiliary monitoring point can be added on the posterolateral thigh. This point is located approximately 6 cm vertically downwards from the midpoint of the line connecting the ischial tuberosity and the greater trochanter. Here, the muscle fiber direction is clearly defined, effectively capturing the concentrated discharge of the lower gluteus maximus during powerful hip extension.

[0079] The gracilis muscle, as a representative of the adductor longus group with significant function and a superficial location, is an ideal choice for monitoring hip adduction and assisted knee flexion. Electrodes are precisely placed on the medial thigh, about 10 cm above the medial aspect of the knee joint, on the longitudinal cord-like muscle belly that can be felt at the posterior border of the sartorius muscle. The signal at this location is clear and can effectively reflect the adduction stability during the gait support phase and the intention of leg folding when sitting.

[0080] Electrode placement locations for calf and foot muscle groups:

[0081] The calf and foot muscles are the final power output and landing stability guarantee for lower limb movement. They dominate the plantar flexion and dorsiflexion of the ankle joint, maintain the arch shape, and directly determine the propulsive force of walking, the smoothness of landing, and the stability of standing. The medial and lateral heads of the gastrocnemius are the core power-generating muscles of the ankle joint plantar flexion, providing forward propulsive force during the walking propulsion phase, heel raise, and squatting push-off phase, and are the final power source for the lower limb push-off force. The soleus muscle, as an endurance stabilizing muscle, together with the gastrocnemius to form the triceps surae, maintains body balance during the standing and mid-walking support phases, prevents the body from leaning forward, and is key to ensuring static and dynamic stability. The tibialis anterior is the only dominant muscle of ankle dorsiflexion, controlling the dorsiflexion of the foot during the heel lift-off and heel strike phases of walking, effectively preventing the foot from dragging and ensuring gait smoothness. The peroneus longus dominates foot eversion and arch stability, maintaining the arch shape throughout the gait support phase, distributing plantar pressure, avoiding foot inversion injury, and assisting in ankle plantar flexion. The coordinated action of the calf and foot muscles enables propulsion, landing cushioning, and postural stability of lower limb movements, and is the core of lower limb stability and propulsion.

[0082] The medial and lateral heads of the gastrocnemius muscle are the primary muscle groups responsible for plantar flexion (heel raise) and propulsion during walking. The medial head electrode is located one palm's width (approximately 10-12 cm) below and behind the medial condyle of the tibia, at the highest point of the prominent bulge on the posteromedial aspect of the lower leg. The lateral head electrode is located one palm's width (approximately 10-12 cm) below and behind the head of the fibula, at the highest point of the prominent bulge on the posterolateral aspect of the lower leg. Signals from these two points directly determine the magnitude and speed of propulsion.

[0083] The soleus muscle is a key endurance muscle for maintaining standing posture and mid-stability during walking, and its signals are crucial for preventing falls. The electrode is placed at the junction of the middle and lower third of the posterior calf, about 3 cm below the lower edge of the gastrocnemius muscle belly and lateral to the medial border of the Achilles tendon, at a deep tender point. This is the thickest part of the soleus muscle, and the area where the signal is most easily captured by the surface electrode.

[0084] The tibialis anterior is the sole dominant muscle controlling dorsiflexion of the foot, enabling heel strike and preventing dragging during the swing phase. Electrodes are positioned longitudinally on the lateral side of the tibia, approximately one palm's width (10-12 cm) below the tibial tuberosity and about 2 cm lateral to the anterior tibial crest, at the most prominent point of the muscle belly. This location provides a strong and specific signal, directly corresponding to the dorsiflexion intention of the ankle joint.

[0085] The peroneus longus muscle plays a crucial role in maintaining foot arch stability, controlling foot eversion, and assisting plantar flexion. Electrodes are obliquely positioned along its course in the upper-middle lateral segment of the lower leg, approximately 6-8 cm below the fibular head, and obliquely downward and backward towards the lateral aspect of the calcaneus. This direction aligns with the direction of the muscle fibers, optimizing the acquisition of its stabilizing signals throughout the entire gait support phase, from heel strike to toe liftoff.

[0086] Data preprocessing:

[0087] To achieve high-fidelity processing of sEMG signals, this invention employs a combined noise reduction strategy using a Butterworth fourth-order bandpass filter and a Butterworth notch filter. The bandpass filter's passband range is set to 20Hz-450Hz, with passband ripple controlled within ±1dB and stopband attenuation greater than 40dB. This effectively suppresses low-frequency baseline drift (<10Hz) and high-frequency noise (500Hz) while preserving the core frequency band of muscle activity. For power frequency interference and its harmonic components, a multi-stage Butterworth notch filter is designed with center frequencies set at 50Hz, 100Hz, and 150Hz, respectively. Each stage has a bandwidth of 2Hz and a quality factor (Q) of 25. Bidirectional filtering (zero phase offset) eliminates signal time-domain distortion.

[0088] The amplitude and frequency characteristics of a Butterworth filter satisfy the following relationship: ;

[0089] In the formula, n is the filter order. The cutoff frequency, For frequency.

[0090] This invention uses the Python SciPy signal processing library to implement a Butterworth filter. The specific parameter settings are shown in Table 1 below: Table 1 Butterworth Filter Parameter Setting Data Table

[0091] Figures 1 to 4 The comparison of the time-domain waveforms of the sEMG signals before and after denoising is shown. The original signal exhibits significant baseline drift and power frequency interference, causing substantial fluctuations in signal amplitude. After Butterworth bandpass filtering and notch filtering, baseline drift and power frequency interference are effectively suppressed, and key electromyographic features are clearly visible. The comparison demonstrates that the denoised signal effectively removes high and low frequency noise while retaining the main frequency information. Furthermore, the goniometer signal is processed using a 10Hz low-pass filter (fourth-order Butterworth).

[0092] Specifically, step S2 includes the following process:

[0093] Electromyography feature extraction:

[0094] Feature extraction aims to extract useful information from sEMG signals and remove unwanted electromyographic components and interference. In electromyographic signal analysis, features are generally classified into three main categories: time-domain, frequency-domain, and time-frequency features. This invention focuses on the analysis of time-domain and frequency-domain features, primarily based on the following considerations: While time-frequency features (such as wavelet coefficients) can reflect the local time-frequency characteristics of a signal, their high-dimensionality requires dimensionality reduction methods before they can be input into a classifier. This process not only increases computational complexity but may also introduce information loss. In contrast, time-domain and frequency-domain features can directly characterize the amplitude modulation characteristics and spectral energy distribution of a signal and possess clear physiological significance.

[0095] Temporal feature extraction:

[0096] Time-domain features are typically fast and easy to implement because they do not require any transformations and are calculated based on the original EMG time series. Time-domain features have been widely used in medical and engineering research and practice.

[0097] Mean Absolute Value (MAV): This describes the average amplitude characteristics of an electromyographic signal, characterizing the signal intensity. Due to its simplicity of calculation, it is widely used in signal feature extraction. Its calculation formula is as follows: ;

[0098] in, This represents the amplitude of the EMG signal at the i-th sampling point, and N is the total number of sampling points for the signal.

[0099] Root Mean Square (RMS): This is a fundamental measure of signal amplitude, characterizing the energy of the signal in the horizontal direction. It is closely related to the signal power and can also be referred to as the effective value of the electromyographic signal within a given period. Its mathematical formula is as follows: .

[0100] Variance (VAR) is a measure of signal power, measuring the distance between each sample point and the sample mean. It describes the dispersion of data. Typically, variance is defined as the average of the squared deviations of a variable from its mean. However, since the mean of an EMG signal is close to zero, the variance of an EMG signal can be simplified to the average of the squares of the signal amplitude. The calculation formula is as follows: ;

[0101] Waveform length (WL) is a metric for measuring the complexity of EMG signals. It is defined as the cumulative length of the EMG waveform over a time period. WL reflects the degree of signal variation over time, thus providing a quantitative indicator of the complexity of muscle activity. Specifically, WL can be calculated using the following formula: ;

[0102] Zero crossing (ZC): A metric for EMG signal frequency information defined in the time domain. It represents the number of times the EMG signal amplitude crosses the zero amplitude level. To reduce the impact of low voltage fluctuations or background noise, a threshold condition is usually set. The formula for calculating ZC is as follows: ; ;

[0103] Autoregressive coefficients (AR): An autoregressive (AR) model is a predictive model that describes each sample of an EMG signal as a linear combination of the previous few samples plus a white noise error term. Specifically, the model can be represented as: ;

[0104] in, This is the current sample. These are the autoregressive coefficients, and p is the order of the model. This is the white noise error term. Autoregressive coefficients. It is used as a feature vector because it can capture the dynamic characteristics of a signal.

[0105] Integrated electromyography (EMG) value This quantifies the total energy of muscle activation, reflecting the intensity of force exertion. Its value is the integral of the absolute value of the electromyographic signal within a time window. ;

[0106] in, It is the electromyographic signal voltage value at time t (unit: μV); These are the start and end times of the analysis time window (in seconds); It is the absolute integral value of the electromyographic signal within the time window (unit: μV·s), reflecting the total activation of the muscle.

[0107] Frequency domain feature extraction:

[0108] Compared to time-domain analysis, frequency-domain analysis is slightly more complex in its computation, but this does not mean that frequency-domain analysis is inferior to electromyography (EMG) signal analysis. On the contrary, increasing research results show that muscle force changes during limb movement, and the time-domain characteristics of EMG signals vary greatly, making them less stable than those in the frequency domain. Power spectral density (PSD) is an important indicator in frequency-domain analysis. PSD is defined as the Fourier transform of the autocorrelation function of the EMG signal. Based on PSD, various statistical properties can be calculated, among which the mean frequency and median frequency are two widely used variables.

[0109] Mean frequency (MNF): This is the weighted average of the EMG signal power spectrum. It is calculated by summing the products of the EMG power spectrum and the frequency, and dividing by the sum of the power spectrum intensities. MNF reflects the frequency characteristics of the signal and is commonly used to assess muscle fatigue and motor unit recruitment. The specific calculation formula is as follows: ;

[0110] in, This represents the frequency of the j-th frequency bin. Let M represent the EMG power spectrum of the j-th frequency bin, where M is the length of the frequency bin.

[0111] Median frequency (MDF): As a common spectral characteristic parameter, it reveals the central tendency of energy distribution in electromyography (EMG) signals and exhibits strong resistance to noise and aliasing interference. Mathematically, MDF can be expressed as: ;

[0112] The median frequency is the frequency point at which 50% of the energy of the power spectrum is accumulated after the Fourier transform of the signal. ;

[0113] in, It is the power spectral density value at frequency f; It is the maximum frequency of the analysis band (take the effective bandwidth of the electromyography signal as 500 Hz);

[0114] The power spectral density (PSD) was calculated piecewise using the Welch method to identify high-frequency noise (>500 Hz) and low-frequency motion traces (<20 Hz).

[0115] ;

[0116] in, It is the first Fourier transform results of segmental electromyographic signals; It is the number of segments (dividing the signal into segments). (Segment, overlapping 50%).

[0117] Specifically, step S3 includes the following process:

[0118] This invention constructs a concise architecture of "multi-scale feature extraction - dynamic channel calibration - lightweight classification" and proposes a lightweight multi-scale convolutional attention network (LMSACNN) adapted for lower limb motion pattern recognition based on multi-channel surface electromyography (sEMG) signals. The network extracts spatiotemporal features of EMG signals through three stacked lightweight multi-scale attention modules (LMSAM), and combines them with a classifier of "1×1 convolution + batch normalization + PReLU activation" to complete motion pattern recognition through a Softmax function, balancing recognition accuracy and deployment efficiency to meet the needs of exoskeleton embedding.

[0119] Adaptive one-dimensional depthwise separable convolution (A-1DDSC):

[0120] This invention addresses the problem of insufficient adaptation of traditional fixed-kernel convolution to the frequency characteristics of electromyographic signals under different motion modes. It innovatively introduces a dynamic convolution kernel selection mechanism, adaptively matching 3rd, 5th, or 7th order convolution kernels based on the average frequency of the current electromyographic signal. This effectively controls computational load while accurately capturing temporal features at different scales. The formula is as follows: ;

[0121] in, (L is the signal time step, and C is the number of electrode channels). The average frequency of the current electromyographic signal is obtained by real-time calculation of the signal spectrum characteristics. Spatial convolution is performed independently for each input channel to focus on local temporal feature extraction; pointwise convolution is performed using a 1×1 convolution kernel to achieve cross-channel feature fusion, and the final output is... Maintain consistency in input and output dimensions to ensure the continuity of processing in subsequent modules.

[0122] Dynamic Efficient Channel Attention (D-ECA):

[0123] To address the issues of susceptibility to noise interference and significant signal quality variations across different channels in electromyography (EMG) signals, a signal-to-noise ratio (SNR) feedback adjustment mechanism is integrated into the traditional ECA attention mechanism. This dynamically adjusts the weight allocation of each channel, ensuring the model prioritizes high-SNR, effective signal channels while suppressing the influence of noise-interfering channels. The core formula is as follows: ; ; ;

[0124] In the formula, γ=2 and b=1 are the optimal empirical parameters verified by experiments, ensuring that the kernel size is odd to improve the symmetry of feature extraction; SNR(X) is the real-time signal-to-noise ratio of the electromyography signal in each channel, and SNRmax=40dB is the preset normalization threshold to avoid excessive imbalance in weight distribution; σ is the Sigmoid activation function, which maps the weight values ​​to the [0,1] interval. The dynamically generated channel weight vector, along with the original feature matrix. Element-by-element multiplication ( After that, the enhanced feature matrix is ​​obtained. This significantly improves the recognizability of effective features.

[0125] Lightweight Multiscale Attention Module (LMSAM):

[0126] To balance multi-scale feature extraction and anti-interference capabilities, this invention employs a dual-path parallel structure, simplifying model complexity while ensuring the integrity of feature representation. The complete module formula is as follows: ;

[0127] The first path passes through (3rd-order convolution kernel) focuses on extracting short-term temporal features of the signal, capturing transient changes in muscle activation; the second path passes through Max pooling (kernel size 3, stride 1) is performed to effectively filter random noise in the signal and enhance the model's robustness to abnormal perturbations. The outputs of the two paths are concatenated by the Concat function to achieve channel dimension concatenation. After the channel weights are dynamically calibrated by the D-ECA module, the number of channels is compressed to C′ (C′ < C) by pointwise convolution. The final output is a module feature that integrates multi-scale features and anti-interference information. This reduces the amount of subsequent computation while retaining key feature information.

[0128] The model training settings are as follows:

[0129] Dataset selection and partitioning:

[0130] The ENABL3S public dataset was selected, covering 10 subjects of different ages and body types, including 7 lower limb movement patterns such as walking on flat ground and sitting up. For each pattern, 500 sets of 14-channel electromyography (EMG) signals were collected per subject (sampling frequency 1000Hz, single sample duration 250ms). A "cross-individual non-overlapping" partitioning strategy was adopted, dividing the dataset into a training set (7 subjects, 3500 sets), a validation set (1 subject, 500 sets), and a test set (2 subjects, 1000 sets) in a 7:1.5:1.5 ratio, ensuring no data overlap between the same subjects and directly verifying cross-subject generalization ability.

[0131] Loss function design:

[0132] For multi-class classification tasks involving lower limb movement patterns, the categorical cross-entropy loss function is used to optimize model parameters. This function effectively measures the difference between the model's predicted probability distribution and the true label distribution, guiding the model to quickly learn the feature discrimination of different movement patterns. The core formula is: ;

[0133] In the formula, N is the number of training samples per batch, and M is the number of motion pattern categories (M=7). The true label for the i-th sample in the c-th motion pattern is (using one-hot encoding, i.e., the correct class label is 1, and the other classes are 0). This function predicts the probability that the i-th sample belongs to the c-th motion pattern. The loss function amplifies the deviation between the predicted probability and the true label through logarithmic operations, allowing the model to prioritize correcting samples with larger prediction errors during training, thus ensuring a balanced recognition accuracy across different motion patterns.

[0134] like Figure 5 As shown, the LMSACNN model proposed in this invention performs excellently in the cross-subject motion recognition task on the ENABL3S dataset, with an average recognition accuracy of 93.07% across all subjects, fully validating the model's adaptability to electromyographic signals from different individuals. Specifically, there are reasonable differences in recognition accuracy among the subjects: Subject S2 achieved the highest accuracy at 95.74%, which is closely related to the stable amplitude of its electromyographic signal and low noise interference; Subject S9 achieved the lowest accuracy at 90.96%, mainly because this subject had weaker muscle strength, lower electromyographic signal amplitude, and slight irregularities in its movements.

[0135] Motion recognition performance based on sEMG signals is easily affected by individual characteristics of the subjects. Different subjects have natural differences in muscle morphology, force exertion habits, and skin impedance, resulting in differences in the amplitude, frequency distribution, and stability of electromyographic signals. Therefore, a fluctuation of 3-5 percentage points in recognition accuracy between individuals is expected. Even so, the recognition accuracy of all subjects remained above 90%, significantly higher than the lowest individual accuracy of the comparison models (82.1% for traditional CNN and 83.5% for conventional ECA-Net). This further demonstrates that the LMSACNN model effectively reduces the negative impact of individual differences through multi-scale feature extraction and dynamic channel calibration mechanisms, and has stable cross-subject generalization ability.

[0136] Specifically, step S4 includes the following process:

[0137] Traditional joint angle prediction models struggle to effectively capture the dynamic spatial correlations and temporal evolution patterns of multi-channel electromyography (sEMG) signals, and are susceptible to interference from redundant information during feature processing, resulting in limited prediction accuracy. This project proposes a channel attention-enhanced LSTM model (CAELSTM), which, through a compact three-stage architecture of "feature extraction-feature refinement-prediction output," accurately captures spatiotemporal dynamic patterns from multi-channel sEMG signals, achieving high-precision continuous prediction of hip, knee, and ankle joint angles, and providing reliable target angle input for exoskeleton collaborative control.

[0138] Feature extraction stage:

[0139] The time-frequency domain electromyography features (including time-domain MAV, RMS, iEMG and frequency-domain PSD, MNF, etc.) extracted using 20 consecutive sliding windows (window length 250ms, step size 20ms) are used to construct the sEMG feature sequence matrix S as the model input, with the expression: ;

[0140] in, This represents the feature vector composed of M features extracted from the w-th window, ensuring the integrity of time series information and the correlation between features.

[0141] CNN embedding layer:

[0142] A one-dimensional convolutional neural network (1D CNN) is used as the embedding layer to transform the original 14-channel time-frequency domain feature sequence matrix. Mapped to a 32-dimensional high-dimensional feature space, the convolution operation formula is: ;

[0143] In the formula, h[t] is the value of the output feature map at position t, w[k] is the weight of the convolution kernel, and S[t+k] is the corresponding position value of the input sequence. This layer makes full use of the temporal locality and cross-channel interaction characteristics of sEMG signals, captures local muscle activation patterns through a weight sharing mechanism, and integrates complementary information from multiple sources to provide high-quality abstract features for subsequent temporal modeling.

[0144] LSTM layer:

[0145] The long-term temporal dependence of electromyographic features is learned through an LSTM network, and an input gate is introduced ( Forgotten Gate ) and output gate ( Three types of gated logic units optimize feature memory and updating. The core formula is as follows: ; ; ;

[0146] In the formula, σ represents the sigmoid activation function, whose output range is (0, 1). It is by and The constructed vector, weight matrix , and These correspond to the forget gate, input gate, and output gate, respectively. Similarly, the bias vector... , and These are related to the forget gate, input gate, and output gate, respectively.

[0147] The candidate cell state is calculated before being controlled by the input gate. It represents the potential new information to be added to the cell state, as shown in the following formula: ;

[0148] In the formula, , Here, represents the weight matrix and bias term of the memory cell, respectively, and tanh is the hyperbolic tangent activation function. Current cell state. Depend on and candidate cell status ,pass and The formula obtained through joint updating is as follows: ;

[0149] In the formula, This represents the dot product of vectors. Current hidden state. The calculation formula for the output information of the LSTM unit is as follows: ;

[0150] Feature refinement stage:

[0151] This project innovatively designs a hybrid attention enhancement module that integrates fine-grained channel attention (FGCAM) and frequency-domain enhanced channel attention (FDECAM) to collaboratively optimize feature quality from dual perspectives. This architecture overcomes the limitations of traditional channel attention in spatiotemporal coupled analysis, enabling the model to simultaneously capture details of local muscle activation in the temporal domain and patterns of energy distribution in the frequency domain, significantly improving the robustness and discriminability of feature representation.

[0152] Fine-grained channel attention module (FGCAM):

[0153] This project extracts intra-channel and inter-channel information using depthwise convolution and pointwise convolution, respectively. It then combines global max pooling and average pooling to generate two-dimensional channel descriptors. Accurate channel weights are obtained through dynamic fusion and correlation calculation, as shown in the following formula: ;

[0154] ε is a trainable fusion parameter (initial value 0.5). and These are the channel descriptors obtained by global max pooling and average pooling, respectively. This formula dynamically adjusts the weights of the two pooling features to retain key information about muscle activation peaks while also considering the overall activation trend, generating a more comprehensive channel representation.

[0155] To further enable effective interaction between global and local channels, a vector cross product operation is used to capture channel correlations at different granularities: ;

[0156] In the formula, M is the correlation matrix. Then, summation is performed along the row and column dimensions to obtain vector representations of different dimensions, and this summation is repeated to obtain the final weight vector for allocating channel weights. The calculation formula is as follows: ;

[0157] In the formula, C represents the number of channels. The channel interaction operations at different granularities highlight relevant features, downplay unimportant feature channels, and more finely allocate weights between different channels. Finally, the resulting weight vector w is multiplied by the input feature map F, as shown below: ;

[0158] In the formula, This indicates element-wise multiplication. The final output feature map is represented by a global-local channel interaction mechanism that enhances key muscle group features, suppresses redundant interference, and improves the interpretability of the feature space.

[0159] Domain Enhanced Channel Attention Module (FDECAM):

[0160] The time-domain signal is transformed to the frequency domain using one-dimensional discrete cosine transform, extracting spectral features and reducing high-frequency noise interference. The core formula is as follows: ;

[0161] In the formula, The length of a single-channel signal is l, and l is the frequency component index (0 to l). -1), These are the sampling points for the time-domain signal. This formula decomposes the time-domain muscle activation signal into different frequency components using discrete cosine transform. The low-to-mid frequency components (20-200Hz) correspond to the effective signal of muscle contraction, while the high-frequency components (>200Hz) are mostly noise. Subsequently, frequency domain attention is used to focus on the effective frequency features to improve the model's anti-interference ability.

[0162] Predicting output stage:

[0163] By aggregating temporal feature contextual information through a Historical Feature Fusion (HFFM) module, a temporal pattern attention network is introduced to dynamically capture the impact of key historical moments on the current perspective prediction. The core formula is as follows: ;

[0164] In the formula, H is the hidden state matrix output by the LSTM layer (dimension L × hidden_size), w is the history window length (set to 10 in experiments), and C is the convolution kernel matrix (dimension k × hidden_size, k is the number of temporal patterns). This formula extracts the time-invariant patterns in the hidden state through the convolution kernel, generating a temporal feature matrix HC (dimension L × k).

[0165] At the same time, the following evaluation function f is also used to calculate the correlation: ;

[0166] In the formula, This represents the i-th row of the time feature matrix (corresponding to a time pattern). Hide the current state. This is the attention weight matrix. The correlation score between the current state and each historical time pattern is calculated using vector inner product, quantifying the influence of different historical moments on the current prediction.

[0167] ;

[0168] Subsequently, The row vectors are passed through Weighting to obtain a vector : ;

[0169] Finally, and By combining these, we obtain the final weighted vector. : ;

[0170] in, , , , .

[0171] The feature vectors from HFFM are input into the final fully connected layer, which uses the PReLU activation function. Finally, the output of this layer predicts the corresponding joint angle.

[0172] Evaluation of prediction results:

[0173] The root mean square error (RMSE) is used to evaluate regression performance, as shown in the formula: ;

[0174] In the formula, n is the length of the sampling sequence. The error between the predicted angle and the actual angle at each time point is represented by this indicator. This indicator can effectively reflect the overall deviation between the predicted value and the actual value. The smaller the value, the higher the prediction accuracy.

[0175] The experiment used sEMG signals from the subjects' walking movements and compared them with the actual joint angle values ​​for validation. The RMSE results of the electromyography model are shown in Table 2 below: Table 2. RMSE Results Data of Electromyography Model

[0176] The results showed that the average RMSE for hip joint angle prediction was 7.94, and for knee joint it was 9.31. Subject S3's predicted angles showed good fit with the actual angles, with the hip joint prediction curve perfectly matching the actual curve trend. Knee joint angle errors were concentrated in the motion switching phase (<10°), meeting the high-precision requirements of exoskeleton collaborative control. The RMSE results for different subjects (S1 hip 7.83, S5 knee 8.82, etc.) indicate that the model maintains stable predictive performance across individual scenarios, providing technical support for the practical application of exoskeletons for the elderly.

[0177] Specifically, step S5 includes the following process:

[0178] Coordinate system construction and lower limb position calculation:

[0179] like Figure 6 As shown, the origin of the carrier coordinate system is the centroid of the IMU (Inertial Measurement Unit), X s The axis points in the direction of IMU movement, Z. s The axis is perpendicular to X. s The axis is pointing downwards, Y s The axis is perpendicular to X. s Z s In a plane, the three axes satisfy the right-hand rule. X s Y s Z sThe angles generated by the axis rotation are represented by Euler angles: pitch, yaw, and roll. In this invention, the coordinate system of each inertial measurement unit and lower limb joint is based on the carrier coordinate system.

[0180] Lower limb movement modeling:

[0181] From a mechanical perspective, the lower limb structure is a transmission structure, similar to a linkage mechanism. The thigh and lower leg can be viewed as rod-like components in a linkage mechanism, rotating under the constraints of the joints. Therefore, the joints of the lower limbs are kinematic pairs connecting components. Figure 7 It is a simplified lower limb linkage model, consisting of two rod-shaped components (thigh and calf), three rotational joints (hip, knee, and ankle), and the foot.

[0182] After determining the simplified model of the lower limbs, it is necessary to establish a corresponding DH coordinate system at the joints according to the DH coordinate system rules so that position calculations can be performed along the coordinate system. The spatial position of the lower limbs is a relative positional relationship, and the purpose of establishing the kinematic model of the lower limbs is to solve the position of the lower limb ends relative to the human torso. Therefore, this invention establishes a reference coordinate system at the center of the pelvis, level with the acetabulum, with the direction consistent with the geographic coordinate system.

[0183] Lower limb position calculation based on positive kinematics:

[0184] Taking point D as an example, the coordinates of point D in the lower limb are solved using forward kinematics. The specific steps are as follows.

[0185] (1) Assume the coordinates of the metatarsal node D in coordinate system {D} are: Since D is the origin of the coordinate system, According to the theory of forward kinematics, the coordinates of point D in coordinate system {C} are: .

[0186] (1);

[0187] Let {C} be the homogeneous transformation matrix from coordinate system {C} to coordinate system {D}, derived from the rotation matrix. Translation matrix Multiplying them together yields the result. Where... The orientation of coordinate system {D} is determined by the three-axis rotation angle from coordinate system {C} to coordinate system {D}. The direction of coordinate system {D} is consistent with the carrier coordinate system of the IMU worn on the foot, and the direction of coordinate system {C} is consistent with the carrier coordinate system of the IMU worn on the lower leg. Assume the attitude angle measurements of the IMUs on the lower leg and foot are respectively... , Then the three-axis attitude angles from coordinate system {C} to coordinate system {D} are: (2);

[0188] The translation distance from coordinate system {C} to coordinate system {D} along the Y-axis is: Therefore, the translation transformation matrix = Substituting the above results into the equation yields the following result: .

[0189] (2) Next, solve for the coordinates of point D in coordinate system {B}. The process is similar to step (1). Assume the attitude angle measurement of the IMU worn on the thigh is... The formulas involved are: (3);

[0190] (3) Find the coordinates of point D in coordinate system {A}. The calculation formulas are: (4);

[0191] Since coordinate system {A} is a fixed coordinate system with the same orientation as the geographic coordinate system, the rotation angle from coordinate system {B} to coordinate system {A} is consistent with the attitude angle measured by the IMU at the thigh.

[0192] (4) Find the coordinates of point D in coordinate system {O}. Both coordinate systems {O} and {A} are fixed coordinate systems, with their orientation consistent with the geographic coordinate system, and there is no rotation relationship. Translation transformation matrix == .

[0193] (5) By multiplying the coordinate system transformation matrices of each step in equations (1)-(4), the final result can be obtained: (5);

[0194] This refers to the coordinates of point D in the coordinate system {O}, which is the spatial position of the lower limb end relative to the human torso.

[0195] To verify the accuracy of the lower limb position calculation, this invention designed "leg raising" and "leg swinging" experiments to test the effectiveness of the calculation method.

[0196] Taking point D as an example, the experiment was designed with four fixed trajectories. The subjects controlled their lower limbs to move smoothly along the designed trajectories. Then, the trajectories obtained by the lower limb position calculation method were compared with the preset trajectories, and the error between the two was calculated to complete the verification.

[0197] (1) Horizontal trajectory: The trajectory in the horizontal direction, that is, the trajectory formed by moving the lower limb along the Y-axis for 25cm with the hip as the origin. Its starting point is (25cm, 25cm) and the ending point is (50cm, 25cm).

[0198] (2) Vertical trajectory: The trajectory in the vertical direction, that is, the trajectory formed by moving the lower limb along the Z-axis for 25cm with the hip as the origin. Its starting point is (25cm, 25cm) and the ending point is (25cm, 50cm).

[0199] (3) 45° oblique trajectory: The trajectory is made in the 45° oblique direction, that is, with the hip as the origin, the lower limb moves along the 45° direction on the YZ plane. Its starting point is (25cm, 25cm) and the ending point is (35.35cm, 35.35cm).

[0200] (4) Semicircular trajectory: The semicircular trajectory on the YZ plane is the arc movement of the lower limb around the hip with a radius of 65cm. The starting point of the trajectory is (0cm, -65cm) and the ending point is (0cm, 65cm).

[0201] Considering the differences in lower limb length parameters among individuals, this invention uses standard parameters to define the length of each part of the lower limb: half hip width. =18.75cm, thigh length =31.3cm, calf length =23.7cm.

[0202] Each experimental trajectory was repeated three times, and the lower limb position calculation method proposed in this invention can obtain the corresponding trajectory for each experiment. For example... Figure 8 As shown, the present invention compares the calculated trajectory with the preset trajectory on the YZ plane. Although the calculated result is slightly different from the preset value, the trend of the calculated trajectory is basically the same as that of the preset trajectory, indicating that the lower limb model and position calculation method of the present invention are correct.

[0203] In addition, the present invention also provides a lower limb movement intention prediction system based on electromyography-inertial signals, including:

[0204] The surface electromyography signal acquisition module includes multiple electrodes for attaching to multiple muscle groups in the lower limbs to acquire surface electromyography signals.

[0205] The inertial measurement module includes multiple inertial measurement units located in each segment of the lower limbs, used to collect attitude angle data;

[0206] Processor, used to perform the following operations:

[0207] The surface electromyography signal is filtered and preprocessed, and time-domain and / or frequency-domain features are extracted.

[0208] The extracted electromyographic features are input into a lightweight multi-scale convolutional attention network to identify the current lower limb movement pattern;

[0209] The extracted electromyographic features are input into a channel attention-enhanced long short-term memory network to predict the target angles of the hip, knee, and ankle joints.

[0210] Based on the posture angle data, the actual angles of each joint of the lower limb are calculated by modeling in the forward kinematics DH coordinate system and compared with the target angles to obtain the angle deviation;

[0211] Based on the angular deviation and the identified lower limb movement patterns, control commands are generated.

[0212] The actuator module includes multiple actuators located at the joints of the lower limb exoskeleton, which are used to receive control commands and adjust the output torque to achieve follow-alert.

[0213] Based on the above system, the present invention also designs an exoskeleton robot mechanical structure and assistance system.

[0214] like Figure 9 and Figure 10 As shown, the left and right legs of the lower limb exoskeleton assistive robot have identical structures. Both legs include a hip joint assist component 3, a knee joint assist component 5, and an ankle joint assist component 9 as an assistive system. The upper end of the hip joint assist component is fixed to the lumbar fixation device 2, and the lower end is fixed to the thigh connection device 4. The upper end of the knee joint assist component is fixed to the thigh connection device, and the lower end is fixed to the knee joint fixation device 8. The upper end of the knee joint assist component is fixed to the knee joint fixation device, and the lower end is fixed to the ankle joint fixation device 10. The lower limb exoskeleton assistive robot also includes a foot support plate 12.

[0215] The hip joint assist component includes two pneumatic muscles, one end of which is fixed to the thigh connecting device, and the other end is fixed to the knee joint fixation device or the waist fixation device respectively. A hip joint sensor 1, a knee joint sensor 6, a knee joint sensor 7 and an ankle joint sensor 11 are set at the joint of the assist component. When in use, they are located above the main muscles of the wearer's thigh. When the air source adds air, the pneumatic muscles of the thigh are filled with air, the diameter of the pneumatic muscles of the thigh increases and the length shortens, forming a smooth elastic movement that simulates the wearer's muscle activity and helps the wearer's lower limb joints extend or contract, providing the wearer with walking assistance.

[0216] Pneumatic artificial muscles are flexible actuation core components, consisting of an internal elastic tube, an external reinforcing fiber layer, and an end connection mechanism. They achieve radial expansion and axial contraction through internal air pressure changes, converting fluid energy into mechanical tension, thus meeting the actuation requirements of lightweight and high-power-density exoskeletons. Their ideal mechanical property equation is:

[0217] In the formula, To simplify the coefficients, D0 is the original diameter of the muscle, and p is the internal working pressure. Let F be the initial angle between the fiber layer and the axial direction, and F be the output driving force.

[0218] The joint driving torque is related to the muscle output force, the equivalent force arm, and the joint rotation angle. The equivalent force arm varies with the knee joint rotation angle. Dynamic changes, the core relationship can be simplified as follows: ;

[0219] in, For joint driving torque, This refers to the amount of muscle contraction. This is a function representing the characteristics of muscle force (under the same air pressure, the greater the contraction length, the smaller the contraction force; under the same contraction length, the greater the air pressure, the greater the contraction force). For equivalent arm

[0220] Furthermore, by optimizing the driving force output through parallel connection of multiple muscles, the relationship between contraction rate and contraction amount is simplified as follows: ;

[0221] In the formula, h is the shrinkage rate, and i is the transmission ratio. The rated length is for the muscles. This selection meets the assistance needs of the hip, knee, and ankle joints, and the maximum output force and stroke are adapted to the biomechanical requirements of mountaineering for the elderly.

[0222] The control architecture of the lower limb exoskeleton assistive robot adopts a two-layer structure design, such as... Figure 11As shown, the upper layer is the biosignal acquisition and processing module, which acquires motion parameters of each segment of the lower limb in real time through an inertial measurement unit and identifies key gait events through feature extraction algorithms. The lower layer is the intelligent decision-making and execution module, which generates time-varying air pressure commands based on gait phase prediction results and preset assistance modes. The air pressure regulation system achieves precise force control of the muscle actuators through a PID algorithm, ultimately outputting an auxiliary torque that conforms to biomechanical characteristics. This architecture realizes a fully closed-loop control from motion perception to assistance execution, with assistance timing accuracy of ±3% of the gait cycle. To determine the joint output torque characteristics of the lower limb exoskeleton assistive robot, a three-dimensional coupling relationship model of air pressure-angle-torque needs to be established. This model obtains basic data through system calibration experiments: measuring the actuator output force under different air pressure and joint angle conditions, and calculating the corresponding joint torque values ​​by combining the geometric parameters of the transmission mechanism. Specifically, tension sensors are used to record rope tension, and the joint driving torque is derived through torque balance equations based on structural parameters such as pulley radius and lever arm length. This calibration method can fully characterize the mechanical performance of the system across its entire operating range, providing accurate torque estimation basis for the control algorithm.

[0223] System experiments were conducted to obtain data on the coupling characteristics of air pressure, tension, and angle: Under stable air pressure conditions, the output of the pressure sensor, the measured value of the rope tension, and the joint angle calculated by the inertial unit were recorded simultaneously. After executing the standardized motion protocol, a two-dimensional relationship between tension and angle under single air pressure was established. Then, multiple sets of pressure condition data were integrated, and a three-dimensional parametric relationship model was constructed using a surface fitting method. This model fully characterizes the mechanical transmission characteristics of the system across the entire operating range, providing a mathematical basis for torque prediction and control. The flowchart of the relationship model calibration experiment is shown below. Figure 12 As shown.

[0224] Experimental data analysis yielded three key characteristic curves: the functional relationship between air pressure and stiffness conversion threshold, the variation of peak tension with pressure, and the correspondence between joint angle and rope tension under specific pressure. These fitted curves comprehensively describe the static mechanical characteristics of the system, providing a basis for optimizing control parameters. Figure 13 As shown, the fitted curves of knee joint angle and cable tension under different working pressure conditions are illustrated. Experimental data show that under constant air pressure, cable tension increases monotonically with the increase of joint flexion angle, with a peak tension of 358.5 N measured at 8 bar pressure. Each pressure curve exhibits a distinct initial segment, within which the tension remains close to zero; the corresponding angle threshold is the critical point for stiffness transition. Notably, this critical angle systematically decreases with increasing working pressure, indicating that the system's assist range expands with increasing pressure. This characteristic provides a theoretical basis for dynamically adjusting the assist range.

[0225] The modeling of the lower limb exoskeleton assistive robot and the assistive process in different scenarios are detailed below:

[0226] Standing up: When standing up, the pneumatic muscles on the back of the hip joint inflate and contract, generating a pulling force from the hip fixation device to the thigh connection device. Simultaneously, the pneumatic muscles on the front of the hip joint deflate and relax, pulling the thigh connection device upward and backward to achieve hip extension. While the hip exerts force, the pneumatic muscles on the front of the thigh deflate and relax, while the two pneumatic muscles on the back of the thigh inflate and contract, transmitting support force to the knee joint fixation device and pushing the knee joint fixation device upward and forward. Throughout the standing up process, the two pneumatic muscles on the back of the calf continuously inflate and contract slightly to maintain stable contact between the ankle joint fixation device and the sole of the foot. The pneumatic muscles on the front of the calf deflate and relax throughout the process, adapting to the natural dorsiflexion tendency of the foot as the center of gravity gradually shifts upward, until the body is fully upright. All inflated muscles maintain a moderate contraction state to maintain standing stability.

[0227] Sitting down: When sitting down, the pneumatic muscles on the front of the hip joint inflate and contract, generating a pulling force from the hip fixation device to the thigh connection device. Simultaneously, the pneumatic muscles on the back of the hip joint deflate and relax, pulling the thigh connection device downward and forward to achieve hip flexion. While the hip exerts force, the two pneumatic muscles on the back of the thigh deflate and relax, while the pneumatic muscles on the front of the thigh inflate and contract, transmitting traction assistance to the knee joint fixation device, causing the knee joint fixation device to move downward and backward. Throughout the sitting process, the pneumatic muscles on the front of the calf continuously inflate and contract, maintaining the angle control between the ankle fixation device and the foot to prevent dragging on the ground. The two pneumatic muscles on the back of the calf deflate and relax throughout the process, adapting to the natural landing tendency of the feet as the center of gravity gradually shifts downward, until the body completes the sitting posture. At this point, all inflated muscles return to a relaxed state, maintaining stable support for the sitting posture.

[0228] Walking: In the initial stage of the walking support phase, the pneumatic muscles on the posterior side of the hip joint on the supporting side inflate and contract first, generating a pulling force from the hip fixation device to the thigh connection device. Simultaneously, the pneumatic muscles on the anterior side of the hip joint deflate and relax, pulling the thigh connection device backward to achieve hip extension and provide propulsion for walking. While the hip exerts force, the pneumatic muscles on the anterior side of the thigh on the supporting side inflate and contract, while the two pneumatic muscles on the posterior side of the thigh deflate and relax simultaneously, transmitting support force to the knee joint fixation device and pushing the knee joint fixation device forward to achieve knee extension and maintain lower limb stability. At the same time, the two pneumatic muscles on the posterior side of the calf on the supporting side inflate and contract, while the pneumatic muscles on the anterior side of the calf deflate and relax, achieving plantar flexion through the ankle joint fixation device and providing support for pushing off the ground. As the heel lifts off the ground and the toes are about to leave the ground, the support phase enters its final stage, and the inflated muscles on the supporting side begin to gradually and slowly deflate, with the pulling force slightly decreasing.

[0229] After the support phase ends, the swing phase begins. The original support side switches to the swing side, and the opposite side becomes the new support side, simultaneously initiating the force exertion of the support phase. The pneumatic muscles on the front of the hip joint on the swing side immediately inflate and contract, while the pneumatic muscles on the back of the hip joint deflate and relax, pulling the thigh connecting device forward to achieve hip flexion and swing the leg forward. Simultaneously with the hip swing, the two pneumatic muscles on the back of the thigh on the swing side inflate and contract, while the pneumatic muscles on the front of the thigh deflate and relax, transmitting traction assistance to the knee joint fixation device, which then moves backward to achieve knee flexion, allowing the leg to bend naturally. At the same time, the pneumatic muscles on the front of the calf on the swing side inflate and contract, while the two pneumatic muscles on the back of the calf deflate and relax, achieving dorsiflexion of the foot through the ankle joint fixation device to prevent dragging of the foot until the heel on the swing side touches the ground, ending the swing phase. That side then switches to the new support side, repeating the above support phase movements. Both lower limbs alternately complete the sequential movements of the support and swing phases, achieving continuous and smooth walking.

[0230] Left Turn: When turning left, the right side is the support side and the left side is the swing side. The pneumatic device on the outer side of the back of the thigh on the support side inflates and contracts more, while the pneumatic device on the inner side contracts less and is slightly longer, causing the thigh connecting device to swing slightly to the left. At the same time, the pneumatic device on the outer side of the back of the lower leg on the support side shortens synchronously, while the inner one is slightly longer. Through the ankle joint fixing device, the sole of the foot rotates slightly to the left, forming the support force basis for the left turn. When the left leg on the swing side swings forward, the pneumatic device on the inner side of the back of the thigh contracts more than the outer one, and the pneumatic device on the inner side of the back of the lower leg also shortens synchronously, pulling the knee joint fixing device and ankle joint fixing device to shift to the left, allowing the swing trajectory of the left leg to extend to the left in an arc. By adjusting the length difference of the two pneumatic devices on the back of the thigh and lower leg on both sides, and coordinating with the linkage of the hip, knee and ankle, a smooth left turn is achieved.

[0231] Human-computer coupling simulation and performance verification in multiple scenarios are detailed below:

[0232] Save the human-machine wearable model as Parasolid format, import it into the simulation analysis software, set the MKS unit system, and set the materials of each part of the lower limb exoskeleton assistive robot, as well as the mass and moment of inertia of the human model.

[0233] To specifically observe the human-machine coupling simulation, it is necessary to add actuators to the hip and knee joints of the lower limb exoskeleton assistive robot. Simulation conditions are set up for typical mountainous environments. The hip joint actuator functions and knee joint actuator functions added for flat terrain, stepped terrain, and sloping terrain are shown in Table 3 below: Table 3. Drive Function Data Table for Lower Limb Assistive Exoskeleton Robot

[0234] To evaluate the performance of the thigh of a lower limb exoskeleton-assisted robot, a planar motion simulation model was established, and its gait characteristics under various terrains, including horizontal ground, steps, and slopes, were systematically analyzed. By defining material parameters, applying motion constraints, and configuring driving conditions, the model successfully obtained complete gait cycle data after dynamic solution.

[0235] In a simulation of walking on flat ground, a complete gait cycle lasted 3.3 seconds. Simulation results showed that the system successfully assisted the user in achieving alternating swinging and support of both lower limbs by coordinating the flexion and extension movements of the hip and knee joints. The joint angle time-history curves extracted from the data processing not only revealed the phase difference between the left and right limb movements but also visually demonstrated the effect of the intervention. Specifically, during the movement, the peak flexion of the right knee reached 65.25°, while the peak flexion of the left knee and both hips was around 28°, reflecting the alternation and coordination of the gait.

[0236] When it comes to standing up and sitting down, the device provides precise assistance tailored to the needs of the elderly. The standing cycle is 3.2 seconds. After recognizing the intention of leaning back and extending the hips and knees, the pneumatic muscles on the back of the hip joint and the front of the knee joint contract simultaneously. The knee joint gradually extends from 82° to 1.2° while the hip joint returns from 38° to 14°, compensating for insufficient strength in the quadriceps and gluteus maximus muscles and ensuring a smooth rise.

[0237] The sitting cycle is 3.5 seconds. Control is smooth, and the pneumatic muscles on the front of the hip joint and the back of the knee joint gradually contract to guide the body to slowly sink. The peak hip and knee flexion is about 38° and 82° respectively. The muscles on the front of the ankle joint contract to maintain dorsiflexion of the foot to prevent dragging, reducing muscle load and the risk of imbalance.

[0238] These simulation results under different conditions not only verify that the lower limb assistive exoskeleton robot can assist users in completing daily activities through coordinated joint assistance, but the extracted key kinematic parameters also provide important data support for the evaluation of system performance and further optimization.

[0239] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting lower limb movement intention based on electromyography-inertial signals, characterized in that, Includes the following steps: S1, Signal Acquisition and Preprocessing: Acquire surface electromyography (EMG) signals from multiple muscle groups in the lower limbs, as well as attitude angle data from the lower limb inertial measurement unit, and perform filtering preprocessing on the surface EMG signals; S2, Electromyography Feature Extraction: Extract time-domain and / or frequency-domain features from the preprocessed surface electromyography signal to construct an electromyography feature sequence; S3, Movement Intent Recognition: The extracted electromyographic features are input into a lightweight multi-scale convolutional attention network to identify the current lower limb movement pattern; S4, Joint Angle Prediction: The extracted electromyographic features are input into a channel attention-enhanced long short-term memory network to predict the target angles of the hip, knee, and ankle joints. S5, Position Calculation and Cooperative Assist Control: Based on the posture angle data, the actual angles of each joint of the lower limb are calculated by modeling in the forward kinematics DH coordinate system and compared with the target angle to obtain the angle deviation; based on the angle deviation and the identified lower limb movement pattern, control commands are generated to adjust the output torque of the actuator of the exoskeleton joint.

2. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 1, characterized in that, In step S1, the multiple muscle groups of the lower limbs include: The gluteus maximus, gluteus medius, gluteus minimus, and iliopsoas muscles in the hip region; The quadriceps femoris, hamstrings, lower gluteus maximus, and gracilis muscles of the thigh; The gastrocnemius muscle (medial and lateral heads), soleus muscle, tibialis anterior muscle, and peroneus longus muscle of the lower leg and foot; The electrodes for the surface electromyography (EMG) signal are arranged along the direction of the muscle fibers of each muscle group and cover the muscle belly region of the corresponding muscle group.

3. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 2, characterized in that, In step S1, the filtering preprocessing employs a combined noise reduction strategy, as follows: The Butterworth fourth-order bandpass filter with a passband range of 20Hz-450Hz was used, as well as multi-stage Butterworth notch filters with center frequencies of 50Hz, 100Hz and 150Hz. The amplitude and frequency characteristics of the Butterworth filter satisfy the following relationship: ； In the formula, n is the filter order. The cutoff frequency, For frequency; This is the frequency response function of the Butterworth filter.

4. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 3, characterized in that, In step S2, the time-domain features include at least one of the following: mean absolute value, root mean square, variance, wavelength, zero-crossing rate, autoregressive coefficient, and integrated electromyography value; the frequency-domain features include at least one of the following: mean frequency, median frequency, and power spectral density. The formula for calculating the integrated electromyographic value is as follows: ； In the formula, It is the electromyographic signal voltage value at time t; , These are the start and end times of the analysis time window, respectively. It is the absolute integral value of the electromyographic signal within the time window, used to reflect the total amount of muscle activation; The median frequency (MDF) satisfies: ； In the formula, It is the power spectral density value at frequency f; It is the maximum frequency of the analysis band.

5. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 4, characterized in that, In step S3, the lightweight multi-scale convolutional attention network includes adaptive one-dimensional depthwise separable convolution and a dynamic and efficient channel attention mechanism; The adaptive one-dimensional depthwise separable convolution is based on the average frequency of the input electromyographic features. The adaptive selection of convolutional kernel size, and the corresponding processing steps are as follows: ； ； The original electromyographic feature matrix is ​​input into an adaptive one-dimensional depthwise separable convolution. L is the signal time step, and C is the number of electrode channels. Represents the real number field; This indicates that a spatial convolution operation is performed independently for each input channel; The output features are obtained by adaptive one-dimensional depthwise separable convolution + pointwise convolution; PointwiseConv = pointwise convolution; The characteristic frequency of the electromyographic signal at time t; The dynamic and efficient channel attention mechanism integrates a signal-to-noise ratio feedback adjustment mechanism to dynamically generate channel weights W. The corresponding processing procedure is as follows: ； ； ； in, The sigmoid activation function is used, SNR(X) is the signal-to-noise ratio of each channel, and SNRmax=40dB is the preset normalization threshold; X is the original feature matrix, Y is the enhanced feature matrix; γ=2, b=1; This indicates element-wise multiplication; GlobalAvgPool represents a one-dimensional sliding convolution on the temporal dimension to extract local features and their association with channels, which are then used to generate channel attention weights. represents the size of the one-dimensional convolution kernel; odd indicates taking the nearest odd number from the calculated result.

6. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 5, characterized in that, In step S4, the channel attention-enhanced long short-term memory network includes a feature extraction stage, a feature refinement stage, and a prediction output stage; The feature extraction stage extracts the temporal dependencies of electromyographic features through a one-dimensional convolutional neural network embedding layer and a long short-term memory network layer. The feature refinement stage includes a fine-grained channel attention module and a frequency-domain enhanced channel attention module. The fine-grained channel attention module generates channel weights through depthwise convolution, pointwise convolution, and global pooling. The frequency-domain enhanced channel attention module transforms the time-domain signal to the frequency domain through a one-dimensional discrete cosine transform to extract spectral features. The transformation formula is as follows: ； in, The length of a single-channel signal is given by l, where l is the frequency component index, i.e., from 0 to l. -1; These are time-domain signal sampling points; is the frequency domain component obtained after the one-dimensional discrete cosine transform; i is the sampling point index of the time domain signal; The prediction output stage aggregates temporal context information through a historical feature fusion module and uses a temporal pattern attention network to capture the impact of historical moments on the current angle prediction.

7. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 6, characterized in that, In step S5, the actual angles of each joint of the lower limb are calculated by modeling using the positive kinematics DH coordinate system, which includes the following process: Based on the linkage model of the lower limb and the DH coordinate system established at each joint, the spatial position of the lower limb distal end in the base coordinate system is calculated through homogeneous coordinate transformation; for the metatarsal node D, the corresponding coordinates in the reference coordinate system {O} are... The solution is obtained using the following formula: ； in, Let D be the coordinates of the metatarsal node D in coordinate system {D}; Let be the homogeneous transformation matrix from coordinate system {C} to coordinate system {D}; Let {B} be the homogeneous transformation matrix from coordinate system {B} to coordinate system {C}. Let be the homogeneous transformation matrix from coordinate system {A} to coordinate system {B}; Let be the homogeneous transformation matrix from coordinate system {O} to coordinate system {A}.

8. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 7, characterized in that, The training process of the lightweight multi-scale convolutional attention network employs a strategy of independent partitioning based on the subjects, ensuring that the subjects included in the training set, validation set, and test set do not overlap, and are divided into training set, validation set, and test set according to a preset ratio; the network training is optimized using the cross-entropy loss function, the expression of which is: ； Where N is the number of training samples per batch, and M is the number of motion pattern categories. Let i be the true label of the i-th sample in the c-th motion pattern. To predict the probability that the i-th sample belongs to the c-th motion pattern.

9. The method for predicting lower limb movement intention based on electromyography-inertial signals according to claim 8, characterized in that, In step S5, the lower limb movement pattern includes one or more of standing up, sitting down, walking on flat ground, and turning left; the actuator is a pneumatic muscle actuator, and the relationship between the corresponding driving force F and the internal working pressure p satisfies: ； in, To simplify the coefficients, D0 is the original diameter of the muscle. The initial angle between the fiber layer and the axial direction; The relationship between joint driving torque, muscle output force, equivalent force arm, and joint rotation angle satisfies: ； in, For joint driving torque, This refers to the amount of muscle contraction. For muscle force driving characteristic function, It is an equivalent arm.

10. A lower limb movement intention prediction system based on electromyography-inertial signals, used to implement the lower limb movement intention prediction method based on electromyography-inertial signals as described in any one of claims 1-9, characterized in that, The lower limb movement intention prediction system based on electromyography-inertial signals includes: The surface electromyography signal acquisition module includes multiple electrodes for attaching to multiple muscle groups in the lower limbs to acquire surface electromyography signals. The inertial measurement module includes multiple inertial measurement units located in each segment of the lower limbs, used to collect attitude angle data; Processor, used to perform the following operations: The surface electromyography signal is filtered and preprocessed, and time-domain and / or frequency-domain features are extracted. The extracted electromyographic features are input into a lightweight multi-scale convolutional attention network to identify the current lower limb movement pattern; The extracted electromyographic features are input into a channel attention-enhanced long short-term memory network to predict the target angles of the hip, knee, and ankle joints. Based on the posture angle data, the actual angles of each joint of the lower limb are calculated by modeling in the forward kinematics DH coordinate system and compared with the target angles to obtain the angle deviation; Based on the angular deviation and the identified lower limb movement patterns, control commands are generated. The actuator module includes multiple actuators located at the joints of the lower limb exoskeleton, which are used to receive control commands and adjust the output torque to achieve follow-alert.

Images