Lower limb motion intention recognition method and system based on electromyographic signals

Abstract

The invention belongs to the technical field of signal processing, and particularly relates to a lower limb motion intention recognition method and system based on electromyographic signals. The method comprises the following steps: S1, arranging electromyographic sensors on a hip key muscle group, a knee joint key muscle group and an ankle joint key muscle group in sequence, and collecting electromyographic signals; s2, performing time-frequency-space domain feature fusion and dimension reduction processing on the electromyographic signals acquired at the hip key muscle group and the knee joint key muscle group, and performing corresponding motion intention prediction by using a dynamic weight distribution method and a decoder based on an LSTM-attention mechanism; electromyographic signals collected at key muscle groups of ankle joints are processed through a muscle force distribution matrix, and meanwhile a time-varying weight adjustment strategy and impedance control are adopted for corresponding motion intention prediction; s3, according to the predicted motion intention, the system controls a direct-current brushless motor on rehabilitation training equipment, a fuzzy self-adaptive PID control method is adopted to regulate and control joint motion, and a walking posture is simulated.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and particularly relates to a method and system for recognizing lower limb movement intentions based on electromyographic signals. Background Art

[0002] Surface electromyographic signals are electrical signals accompanying muscle contractions, which can reflect the degree of muscle contraction. By analyzing surface electromyographic signals, the movement state of muscles can also be recognized. Therefore, they are widely used in many applied scientific fields such as medicine, rehabilitation, and sports science.

[0003] With the development of technology, some rehabilitation trainings have introduced preliminary technologies based on electromyographic signals, but there are still many problems in the existing technologies. On the one hand, in terms of movement intention recognition, the existing methods do not extract the features of electromyographic signals comprehensively and deeply enough. Most technologies only simply extract a few time-domain or frequency-domain features of electromyographic signals, and fail to fully explore the rich information contained in electromyographic signals in the time domain, frequency domain, and spatial domain, resulting in low accuracy of movement intention recognition. For example, only the integrated electromyogram value is used to judge the muscle force intensity, while ignoring the features of the signal in terms of frequency distribution, multi-channel spatial correlation, etc., making it difficult to accurately recognize the lower limb movement intentions of patients in complex movement scenarios.

[0004] On the other hand, in the control link of rehabilitation training, the traditional PID control method uses fixed proportional, integral, and differential coefficients, which are difficult to adapt to individual differences of patients and the dynamic changes of the body state during the rehabilitation process. Since the physiological characteristics such as muscle strength and joint range of motion of different patients are different, and during the rehabilitation training process, the muscle state and movement ability of patients will change continuously, the PID controller with fixed parameters cannot adjust the control strategy in real time, and problems such as insufficient control accuracy and slow system response are likely to occur, thus affecting the effect and safety of rehabilitation training. In addition, the existing rehabilitation training systems lack effective coordination between electromyographic signals and the control of rehabilitation devices, and cannot accurately and timely adjust the movement mode and parameters of rehabilitation devices according to the real-time movement intentions of patients, resulting in poor comfort and efficiency of rehabilitation training and unable to meet the personalized and precise rehabilitation needs of patients.

[0005] Therefore, it is very important to design a method and system for recognizing lower limb movement intentions based on electromyographic signals that can effectively overcome the deficiencies of traditional rehabilitation training and improve the safety, comfort, and efficiency of rehabilitation training. Summary of the Invention

[0006] The present invention aims to overcome the problems of insufficient traditional rehabilitation training in the prior art, and provides a method and system for recognizing lower limb movement intentions based on electromyographic signals that can effectively overcome the deficiencies of traditional rehabilitation training and improve the safety, comfort, and efficiency of rehabilitation training.

[0007] To achieve the above-mentioned invention object, the present invention adopts the following technical solutions: A lower limb movement intention recognition method based on electromyographic signals, comprising the following steps: S1, sequentially arranging electromyographic sensors on the key muscle groups of the hip, knee joint, and ankle joint of the patient and collecting electromyographic signals; S2, subjecting the electromyographic signals collected at the key muscle groups of the hip and knee joint of the patient to time-frequency-spatial domain feature fusion and dimensionality reduction processing, and simultaneously using a dynamic weight allocation method and a decoder based on the LSTM-attention mechanism for corresponding movement intention prediction; subjecting the electromyographic signals collected at the key muscle groups of the ankle joint of the patient to a muscle force distribution matrix processing, and simultaneously adopting a time-varying weight adjustment strategy and impedance control for corresponding movement intention prediction; S3, according to the predicted movement intention, the system controls the DC brushless motor on the rehabilitation training device, and adopts a fuzzy adaptive PID control method to achieve the regulation of joint movement and simulate the walking posture of a real person.

[0008] Preferably, in step S1, the key muscle groups of the hip include: gluteus maximus, gluteus medius / gluteus minimus, and iliopsoas; The key muscle groups of the knee joint include: quadriceps femoris, hamstring muscles, and adductor muscles; The key muscle groups of the ankle joint include: triceps surae, tibialis anterior, peroneus longus, and peroneus brevis.

[0009] Preferably, in step S2, the time-frequency-spatial domain feature fusion and dimensionality reduction processing include the following steps: S21, time domain feature extraction: Set the integrated electromyogram value , which is used to quantify the total energy of muscle activation and reflect the force intensity; set The value of is the absolute value integral of the electromyographic signal within the time window, and the specific formula is: ; Among them, represents the electromyographic signal voltage value at time t; is the start time of the analysis time window; is the end time of the analysis time window; Set the root mean square RMS, which is used to reflect the instantaneous intensity of dynamic muscle contraction, and the specific calculation formula is as follows: ; Among them, is the electromyographic signal value at the i-th sampling point; is the number of sampling points within the time window; is the effective amplitude of the EMG signal, representing the instantaneous muscle contraction intensity; S22, Frequency-domain feature extraction: The median frequency is the frequency point at which the power spectrum accumulates 50% of the energy after the Fourier transform of the signal. The specific calculation formula is as follows: ; where, is the power spectral density value at frequency f; is the maximum frequency of the analysis frequency band, specifically taking the effective bandwidth of the EMG signal as 500 Hz; is the median frequency, reflecting the central tendency of the energy distribution of the EMG signal; The power spectral density PSD is calculated by the Welch method to segment the frequency-domain energy distribution, which is used to identify high-frequency noise and low-frequency motion artifacts. The specific calculation formula is as follows: ; where, is the Fourier transform result of the th segment of the EMG signal; is the number of segments, that is, the signal is divided into segments with 50% overlap; the high-frequency noise refers to the noise with a frequency greater than 500 Hz; the low-frequency motion artifact refers to the motion artifact with a frequency less than 20 Hz; S23, Spatial-domain feature extraction: Construct a covariance matrix based on the multi-channel electrode array signal, and at the same time identify the deep muscle activation pattern through matrix eigenvalue decomposition. The specific formula is as follows: ; where, is the multi-channel EMG signal vector at the th time point, with a dimension of , is the number of electrode channels; is the mean vector of the signals of each channel, with a dimension of ; is the number of samples within the time window; is the covariance matrix, with a dimension of , which is used to characterize the spatial correlation of the multi-channel signals; S24, Feature fusion and dimensionality reduction: The feature fusion includes the following process: Multi-domain feature concatenation: Concatenate the time-domain features , , the main-band energy of the frequency-domain features , and the spatial-domain features into a high-dimensional vector : ; Among them, is the integral value of the power spectral density in the frequency band of 50Hz - 150Hz; the spatial domain feature refers to the principal component of the covariance matrix, including the largest eigenvalue of the covariance matrix , the second largest eigenvalue of the covariance matrix , and the principal component direction angle of the covariance matrix ; Normalization processing: Use Z - score standardization to eliminate the differences in different feature dimensions. The specific formula is as follows: ; Among them, is the mean vector of each feature in the training dataset; is the standard deviation vector of each feature in the training dataset; is the standardized feature vector; The dimensionality reduction process is as follows: Locally Linear Embedding: Map high - dimensional features to a low - dimensional space and preserve the topological structure of muscle group co - activation. The specific process is as follows: ; Among them, the low - dimensional space specifically refers to 3 - D; is the weight matrix, indicating the contribution of neighborhood samples to the reconstruction target; is the feature vector of the i - th sample in the high - dimensional feature space; is the feature vector of the j - th sample in the high - dimensional feature space; is the neighborhood set of the i - th sample; is the weight of the j - th neighborhood sample for reconstructing the i - th sample.

[0010] Preferably, in step S2, the dynamic weight assignment method includes the following steps: Calculation of muscle group contribution degree: Dynamically adjust the feature weights according to the task type (hip extension, abduction, hip flexion). The specific formula is as follows: ; Among them, is the signal - to - noise ratio of the - th muscle group; is the weight adjustment factor, used to control the steepness of weight assignment; is the feature weight of the - th muscle group, with a value range of 0 - 1, and the sum of weights ; The task types include hip extension, abduction, and hip flexion.

[0011] Preferably, in step S2, the network structure of the decoder based on the LSTM - attention mechanism is as follows: Input layer, specifically the feature vector after dimensionality reduction ; LSTM layer, specifically 3 layers of hidden units, used to capture the temporal dependencies of EMG signals; Attention layer, used to calculate the importance weights of features at each time step: ; Among them, is the attention weight at the time step; is the bias vector of the attention mechanism; is the attention weight vector; is the weight matrix of the attention mechanism; is the hidden state vector of the LSTM at time step t; Loss function, specifically a loss function that jointly optimizes the angular error and dynamic smoothness , the specific formula is as follows: ; Predicted hip joint angle; Actual hip joint angle; Predicted thigh angular velocity; represents the thigh angular velocity of the predicted thigh angular velocity at time step t-1. is the smoothing factor, used to control the smoothness of the predicted trajectory; T represents the total number of time steps.

[0012] Preferably, in step S2, the muscle force distribution matrix is set as follows: ; : The acting force of the ankle joint moving towards the sole of the foot.

[0013] : The acting force of the ankle joint moving towards the dorsum of the foot.

[0014] : The acting force of the sole of the foot turning inwards.

[0015] : The acting force of the sole of the foot turning outwards.

[0016] Preferably, in step S2, the time-varying weight adjustment strategy is specifically as follows: The setting formula is as follows: ; Among them, represents at At time t, the weight coefficient of the i-th row and j-th column in the muscle force distribution matrix is used to determine the contribution of different muscle activation levels to the output force of the target movement, and is dynamically adjusted as time and signal quality change; is the basic weight; , represents the adaptive gain coefficient, which is used to determine the weight coefficient Sensitivity to changes in signal-to-noise ratio; The larger the value, the more drastic the response of the weight coefficient to the change of the signal-to-noise ratio; conversely, The smaller the value, the smoother the change of the weight coefficient; yes The real-time signal-to-noise ratio of the i-th channel signal at time.

[0017] Preferably, in step S2, the impedance control process is specifically as follows: The control equation is set as follows: ; in, represents the desired joint angle; represents the expected joint angular velocity; is the actual joint angle; is the actual joint angular velocity; represents the time-varying stiffness; represents the damping coefficient; Proportional term , used to quickly correct desired joint angles and actual joint angles Deviation between differential term , used to suppress the oscillation of the system; feedforward term , used to compensate for the torque required for joint movement in advance according to muscle force; The calculation method of the feedforward term is: ; in, is the total number of muscle groups involved in the calculation; Assigning matrix output force to muscle force; is the tendon moment arm; It represents the angle between the muscle tension line and the joint axis.

[0018] Preferably, in step S3, the fuzzy adaptive PID control method has the following specific process: S31, set the improved fuzzy adaptive PID controller, the corresponding control law is expressed as: ; in, is the output of the controller and is used to drive the controlled object; is the angle deviation, corresponding to the desired angle and the actual angle The difference, that is ; The angle deviation , which is used to reflect the current control error of the system and is an important basis for the controller to make adjustments; is the deviation change rate and is used to reflect the change trend of the error over time; , , are all time-varying control parameters. Among them, corresponds to the proportionality, corresponds to the integral, corresponds to the differential coefficient.

[0019] S32. In the fuzzification process, triangular membership functions are used to achieve quantization. The specific formula is as follows: ; Among them, is the center value of the fuzzy subset; is the width of the membership function; is the membership function used to map the angle deviation to the fuzzy subset .

[0020] It is set that the fuzzy rule base adopts the Mamdani-type rule form with two inputs and three outputs: ; Among them, The k th fuzzy rule in the fuzzy rule base; and are the two inputs of the fuzzy controller, is the angle deviation, is the deviation change rate; , , are the three outputs, respectively representing the adjustment amounts of the proportionality, integral, and differential coefficients; , , , , are the corresponding fuzzy subsets respectively; S33. The defuzzification is carried out by the centroid method. The specific setting formula is as follows: ; Among them, The adjustment amount of the proportionality coefficient ; is the number of fuzzy subsets, is , for the th fuzzy subset membership degree; is the th representative value of the fuzzy subset; The centroid method calculates the centroid position of the fuzzy set according to the shape and distribution of the membership function, and takes the centroid position as the precise output after defuzzification; Finally, the adjustment amount of the proportional coefficient is obtained ; According to the process of step S33, and are calculated in the same way; S34. According to the parameter adjustment amounts , and obtained by defuzzification, the parameters of the PID controller are adjusted online; The parameter adjustment mechanism is as follows: ; Among them, , , respectively represent the real-time values of the proportional coefficient, integral coefficient and differential coefficient of the PID controller at time t . As time t changes, these coefficients will be dynamically adjusted according to the operating conditions of the system, so that the controller can better adapt to the characteristics and control requirements of the system, thereby improving the control effect. , and are all the initial values of the PID controller parameters; , , are all adjustment coefficients, which are specifically obtained through experimental calibration; The adjustment coefficients are used to control the amplitude of parameter adjustment and are reasonably set according to the characteristics and control requirements of the system.

[0021] The present invention also provides a lower limb movement intention recognition system based on electromyographic signals, including; An electromyographic signal acquisition module, which is used to arrange electromyographic sensors on the key muscle groups of the hip, knee joint and ankle joint of the patient in sequence and collect electromyographic signals; A motion intention prediction module, which is used to perform time-frequency-spatial domain feature fusion and dimensionality reduction processing on the electromyography signals collected at the hip key muscle groups and knee key muscle groups of the patient. At the same time, a dynamic weight allocation method and a decoder based on the LSTM-attention mechanism are used to perform corresponding motion intention prediction, and the electromyography signals collected at the ankle key muscle groups of the patient are processed through a muscle force distribution matrix. At the same time, a time-varying weight adjustment strategy and impedance control are used to perform corresponding motion intention prediction; A control and regulation module, which is used to control the DC brushless motor on the rehabilitation training device according to the predicted motion intention, and uses a fuzzy adaptive PID control method to realize the regulation of joint motion and simulate the walking posture of a real person.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Accurately predict motion intention: The prior art has insufficient extraction of electromyography signal features. The present invention comprehensively fuses time-frequency-spatial domain features. For example, the integrated electromyography value and root mean square are used to measure muscle force, the frequency characteristics are analyzed through the median frequency, etc., and a covariance matrix is constructed to obtain spatial domain features. At the same time, the dynamic weight allocation is combined with the decoder of the LSTM-attention mechanism, and the weights are adjusted according to the task, accurately focusing on the key features and improving the prediction accuracy.

[0023] (2) Optimize joint motion control: The present invention adopts a muscle force distribution matrix and a time-varying weight adjustment strategy at the ankle joint, and optimizes the weights in real time according to the signal-to-noise ratio, providing a reliable basis for impedance control. The fuzzy adaptive PID control adjusts the parameters in real time according to the system state, which is more flexible and accurate than the traditional PID control. In impedance control, the proportional, derivative, and feedforward terms cooperate to ensure accurate joint motion.

[0024] (3) Improve the quality of rehabilitation training: The present invention can achieve accurate prediction and control, making the movement of the device fit the needs of the patient, reducing the risk of injury and improving safety. The movement is natural and smooth, reducing the discomfort of the patient and enhancing comfort. The movement mode and parameters are adjusted in time, enhancing the pertinence of training and improving the rehabilitation efficiency.

[0025] (4) Facilitate the development of rehabilitation technology: The innovative method of the present invention opens up a new direction for lower limb rehabilitation treatment and provides ideas for subsequent research. It promotes the upgrading of devices such as lower limb exoskeleton rehabilitation robots to be more intelligent and accurate, meets the personalized needs of patients, and promotes the progress of the rehabilitation treatment technology industry. Description of the Drawings

[0026] Figure 1 It is a schematic curve diagram of the power spectral density of different hip muscle groups changing with the electromyography signal frequency in the present invention; Figure 2 It is a schematic curve diagram of the hip joint angle, actual hip joint angle, and predicted thigh angular velocity changes in the present invention; Figure 3 It is a schematic three-dimensional diagram showing the change of feature weights during knee extension in the present invention; Figure 4 It is a schematic curve diagram showing the change of knee joint angle in the present invention; Figure 5 It is a schematic curve diagram showing the change of ankle joint angle in the present invention; Figure 6 It is a schematic curve diagram showing the change of hip joint, knee joint and ankle joint angles after relevant experiments are carried out through the present invention. Detailed implementation manners

[0027] In order to more clearly illustrate the embodiments of the present invention, the specific implementation manners of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings and other implementation manners can be obtained.

[0028] The present invention provides a method for recognizing lower limb movement intention based on electromyographic signals, including the following steps: 1. Arrange electromyographic sensors on the key muscle groups of the hip, knee and ankle joints of the patient in sequence and collect electromyographic signals; 2. Process the electromyographic signals collected at the key muscle groups of the hip and knee joints of the patient through time-frequency-spatial domain feature fusion and dimensionality reduction, and at the same time use the dynamic weight allocation method and the decoder based on the LSTM-attention mechanism to perform corresponding movement intention prediction; process the electromyographic signals collected at the key muscle groups of the ankle joint of the patient through the muscle force distribution matrix, and at the same time adopt the time-varying weight adjustment strategy and impedance control to perform corresponding movement intention prediction; 3. According to the predicted movement intention, the system controls the DC brushless motor on the rehabilitation training device and adopts the fuzzy adaptive PID control method to realize the regulation of joint movement and simulate the walking posture of a real person.

[0029] For step 1, the key muscle groups of the hip include: gluteus maximus, gluteus medius / gluteus minimus and iliopsoas; The key muscle groups of the knee include: quadriceps femoris, hamstring muscles and adductor muscles; The key muscle groups of the ankle include: triceps surae, tibialis anterior, peroneus longus and peroneus brevis.

[0030] Gluteus maximus: The gluteus maximus is one of the main muscle groups that generate force during hip extension. In actions such as rising from a sitting position to a standing position and the backward leg push during running, the gluteus maximus plays a crucial role. The sacrum is located at the lowest end of the spine and is a relatively flat bone. It is in the middle of the upper part of the buttocks. Part of the origin of the gluteus maximus is on its outside; the greater trochanter of the femur is the prominent bony structure at the uppermost part of the outer side of the thigh. Surface electrodes are arranged along the direction of the muscle fibers of the gluteus maximus. The specific operation is to start from the outside of the sacrum and place the electrodes in the direction of the greater trochanter of the femur. Such an arrangement can closely fit the course of the gluteus maximus, thus effectively capturing the strong electromyographic signals generated by the gluteus maximus during hip extension, providing a high-quality data basis for the subsequent analysis of hip extension movement intention.

[0031] Gluteus medius / gluteus minimus: The gluteus medius and gluteus minimus are anatomically close and have similar functions. They are mainly involved in the abduction movement of the hip joint, such as the side leg lift action. The iliac crest is the bony edge above the hip bone and can be felt on both sides below the waist by hand; the greater trochanter is the prominent part at the uppermost part of the outer side of the thigh mentioned above. The midpoint of the line connecting the lower edge of the iliac crest and the greater trochanter is exactly the area where the gluteus medius and gluteus minimus are located. Since these two muscles are relatively deep, it is difficult for ordinary electrodes to collect high-quality electromyographic signals. Therefore, a high-density electrode array is needed for deep electromyographic acquisition. The high-density electrode array can increase the contact area between the electrodes and the muscle tissue and the sampling points, improving the resolution of signal acquisition; in this way, the separation degree of the abduction action signal can be enhanced, and the electrical activity characteristics of the gluteus medius and gluteus minimus during hip joint abduction movement can be more accurately identified, thus providing more accurate information for the calculation of movement intention.

[0032] Iliopsoas muscle: The iliopsoas muscle is formed by the fusion of the iliac muscle and the psoas major muscle, and its main function is hip flexion. The inguinal ligament is a tough connective tissue band connecting the anterior superior iliac spine and the pubic tubercle, located at the junction of the abdomen and the thigh. The anterior superior iliac spine is the bony protrusion at the front end of the iliac crest; the pubic symphysis is the part in the middle of the front of the pelvis where the left and right pubic bones are connected. Electrodes are placed in the area along the line connecting the anterior superior iliac spine to the pubic symphysis inside the inguinal ligament. This position can accurately approach the iliopsoas muscle. However, in this area, the iliopsoas muscle is easily interfered by the activities of the abdominal muscles, resulting in inaccurate electromyographic signals. To solve this problem, combined with the dynamic impedance matching technology, this technology can adjust the impedance matching relationship between the electrodes and the muscles in real time according to the muscle impedance changes of different individuals and different movement states, reducing the cross-interference generated by the activities of the abdominal muscles, so as to ensure that the collected signals mainly come from the iliopsoas muscle and improve the accuracy and reliability of the iliopsoas muscle electromyographic signal acquisition.

[0033] Quadriceps femoris: The quadriceps femoris consists of the rectus femoris, vastus lateralis, and vastus medialis. For the rectus femoris, place the electromyography sensor on the midline of the anterior thigh, at the midpoint of the line connecting the anterior superior iliac spine and the upper edge of the patella. The acquisition position of the vastus lateralis is on the lateral side of the thigh, slightly medial to the iliotibial tract, approximately at the midpoint of the line connecting the greater trochanter and the lateral edge of the patella. For the vastus medialis, place the sensor on the medial side of the thigh, about 2 - 3 cm above the upper medial edge of the patella. These positions can effectively collect the electromyography signals generated by the quadriceps femoris during knee extension movements.

[0034] Hamstring muscles: The hamstring muscles include the semitendinosus, semimembranosus, and biceps femoris. For the semitendinosus, place the electromyography sensor on the medial side of the popliteal crease on the posterior thigh, about 5 cm above the origin of the semitendinosus tendon. The acquisition point of the semimembranosus is slightly medial and deep to the origin of the semitendinosus tendon. Since its position is relatively deep, the sensor needs to be placed as close as possible to obtain effective signals. For the long head of the biceps femoris, place the sensor on the lateral side of the popliteal crease on the posterior thigh, about 4 - 5 cm above the origin of the biceps femoris tendon. These positions can effectively collect the electromyography signals generated by the hamstring muscles during knee flexion and hip extension movements.

[0035] Adductor muscles: For the gracilis, place the electromyography sensor on the medial side of the thigh, about 8 - 10 cm above the medial condyle of the knee. This muscle is superficial at this position, and the signal is easy to collect. The pectineus is located on the upper medial side of the thigh, close to the pubic symphysis. Place the sensor about 3 - 4 cm below and lateral to the pubic tubercle. The adductor longus is below and lateral to the pectineus. Place the sensor slightly below the midpoint of the line connecting the pubic symphysis and the medial epicondyle of the femur. The adductor brevis is relatively deep and difficult to collect directly. Indirect collection can be performed at a position slightly above the acquisition point of the adductor longus, close to the pubis. These positions can effectively collect the electromyography signals generated by the adductor muscles during hip adduction movements.

[0036] Triceps surae: Lateral head of the gastrocnemius: Place the electromyography sensor on the outer side of the posterior calf, approximately at the upper one-third of the line connecting the midpoint of the calf and the lateral malleolus. This position can effectively collect the electromyography signals generated by the lateral head of the gastrocnemius during ankle plantar flexion movements.

[0037] Medial head of the gastrocnemius: Place the sensor on the inner side of the posterior calf, at the corresponding position to the lateral head, that is, at the upper one-third of the line connecting the midpoint of the calf and the medial malleolus. This position can effectively collect the electromyography signals generated by the medial head of the gastrocnemius during ankle plantar flexion movements.

[0038] Soleus: Due to its relatively deep position, place the electromyography sensor in the middle of the posterior calf, about 5 - 7 cm above the origin of the Achilles tendon. Although the soleus is covered by other muscles, relatively stable and effective electromyography signals can be obtained at this position. These signals are of great significance for analyzing the participation degree and movement intention of the soleus during ankle plantar flexion.

[0039] Tibialis anterior muscle: Place the electromyography sensor on the anterior side of the lower leg, about 1-2 cm lateral to the lateral margin of the tibia, and at the midpoint of the line connecting the midpoint of the lower leg and the ankle joint. This position can more efficiently collect the electromyography signals generated by the tibialis anterior muscle during dorsiflexion and inversion of the foot.

[0040] Peroneus longus and peroneus brevis muscles: Peroneus longus muscle: Place the sensor on the lateral side of the lower leg, about 3-4 cm below the head of the fibula. At this position, the electromyography signals generated by the peroneus longus muscle during eversion of the foot and assisting plantar flexion of the ankle joint can be effectively collected.

[0041] Peroneus brevis muscle: Place the sensor about 2-3 cm below the peroneus longus muscle sensor, close to the lateral malleolus. This position can effectively collect the electromyography signals of the peroneus brevis muscle during eversion of the foot and assisting plantar flexion of the ankle joint.

[0042] For step 2, it actually includes the hip-thigh coordinated movement intention resolution process based on electromyography signals, the thigh-calf coordinated movement intention resolution process based on electromyography signals, and the calf-foot coordinated movement intention resolution process based on electromyography signals.

[0043] Among them, the hip-thigh coordinated movement intention resolution process based on electromyography signals is as follows: Feature extraction and movement intention modeling: Time-domain feature extraction: Set the integrated electromyogram value , which is used to quantify the total energy of muscle activation and reflect the force intensity; set The value of is the absolute value integral of the electromyography signal within the time window, and the specific formula is: ; Among them, Represents the electromyography signal voltage value at time t; Is the start time of the analysis time window; Is the end time of the analysis time window; Set the root mean square RMS, which is used to reflect the instantaneous intensity of dynamic muscle contraction. The specific calculation formula is as follows: ; Among them, Is the electromyography signal value at the i-th sampling point; Is the number of sampling points within the time window; Is the effective amplitude of the electromyography signal, which characterizes the instantaneous muscle contraction intensity; Frequency-domain feature extraction: The median frequency is the frequency point where the power spectrum accumulates 50% energy after Fourier transform of the signal. The specific calculation formula is as follows: ; in, is the power spectrum density value at frequency f; To analyze the maximum frequency of the frequency band, the effective bandwidth of the EMG signal is specifically taken as 500 Hz; is the median frequency, reflecting the central tendency of the energy distribution of the electromyographic signal; The power spectral density (PSD) uses the Welch method to calculate the frequency domain energy distribution in segments, which is used to identify high-frequency noise and low-frequency motion edges. The specific calculation formula is as follows: ; in, For the Fourier transform results of segment EMG signals; is the number of segments, which means the signal is divided into segments, with an overlap of 50%; the high-frequency noise refers to noise with a frequency greater than 500 Hz; the low-frequency motion edge refers to motion edge with a frequency less than 20 Hz; Among them, the power spectrum density of different hip muscle groups changes with the frequency of electromyographic signals as shown in the following figure: Figure 1 shown.

[0044] The gluteus maximus has a high power spectral density in the low-frequency range (20Hz-100Hz), consistent with the sustained force of hip extension. The gluteus medius and minimus are prominent in the mid-frequency range (100Hz-300Hz), which is consistent with the demands of hip abduction. The iliopsoas has a distinct energy distribution in the mid- and high-frequency range (200Hz-500Hz), which facilitates hip flexion. These curves provide important evidence for identifying hip movement intentions and controlling rehabilitation training.

[0045] Spatial feature extraction: The covariance matrix is constructed based on the multi-channel electrode array signal, and the deep muscle activation pattern is identified by matrix eigenvalue decomposition. The specific formula is as follows: ; in, For the The multi-channel EMG signal vector at each time point has a dimension of , is the number of electrode channels; is the mean vector of each channel signal, with dimension ; is the number of samples in the time window; is the covariance matrix, with dimension , used to characterize the spatial correlation of multi-channel signals; Feature fusion and dimensionality reduction: The feature fusion includes the following processes: Multi-domain feature splicing: combining time domain features 、 , frequency domain features and the main frequency band energy of, and the spatial domain features are spliced into a high-dimensional vector : ; Among them, is the integral value of the power spectral density in the 50Hz - 150Hz frequency band; the spatial domain features refer to the main components of the covariance matrix, including the largest eigenvalue of the covariance matrix , the second largest eigenvalue of the covariance matrix , the direction angle of the main components of the covariance matrix ; Normalization processing: Z-score standardization is adopted to eliminate the differences in different feature dimensions. The specific formula is as follows: ; Among them, is the mean vector of each feature in the training dataset; is the standard deviation vector of each feature in the training dataset; is the standardized feature vector; The dimensionality reduction process is as follows: Locally linear embedding: Map the high-dimensional features to a low-dimensional space and preserve the topological structure of muscle group co-activation. The specific process is as follows: ; Among them, the low-dimensional space specifically refers to 3 dimensions; is the weight matrix, indicating the contribution of neighborhood samples to the reconstruction target; is the feature vector of the i-th sample in the high-dimensional feature space; is the feature vector of the j-th sample in the high-dimensional feature space; is the neighborhood set of the i-th sample; is the weight of the j-th neighborhood sample for reconstructing the i-th sample.

[0046] In step 2, the dynamic weight allocation method includes the following steps: Calculation of muscle group contribution degree: Dynamically adjust the feature weights according to the task type (hip extension, abduction, hip flexion). The specific formula is as follows: ; Among them, is the signal-to-noise ratio of the -th muscle group; is the weight adjustment factor, used to control the steepness of weight allocation; is the feature weight of the -th muscle group, and the value range is 0 - 1, and the sum of weights ; The task types included hip extension, abduction, and hip flexion.

[0047] The configuration of the re-adjustment factor parameters was determined experimentally as shown in Table 1 below: Table 1 Weight adjustment factor parameter configuration data table

[0048] In step 2, the network structure of the decoder based on the LSTM-attention mechanism is as follows: Input layer, specifically the feature vector after dimensionality reduction ; Long short-term memory (LSTM) layer, specifically 3 layers of hidden units, used to capture the temporal dependencies of EMG signals; Attention layer, used to calculate the feature importance weights at each time step: ; in, is the attention weight of the time step, , used to quantify the contribution of the current time step to the final prediction; is the bias vector of the attention mechanism; is the attention weight vector; is the weight matrix of the attention mechanism; is the hidden state vector of LSTM at time step t; This formula is used to calculate the time step The attention weight First, through the weight matrix The LSTM is used at time step The hidden state vector Mapped to the attention space, with the bias vector After addition, through the hyperbolic tangent function Perform nonlinear transformation to obtain a set of intermediate feature representations. Then, the attention weight vector These intermediate features are weighted and summed to obtain a scalar value. Finally, the scalar value is passed through The function is normalized to obtain an attention weight between 0 and 1 This calculation method can effectively quantify the importance of each time step feature to the final prediction result, allowing the model to focus on key time series features and improve the accuracy of the prediction.

[0049] Loss function, specifically a loss function that jointly optimizes angle error and dynamic smoothness , the specific formula is as follows: ; Predicted hip joint angle; Actual hip joint angle; Predicted thigh angular velocity; Denotes the thigh angular velocity of the predicted thigh angular velocity at time step t-1. Is a smoothing factor used to control the smoothness of the predicted trajectory; T Represents the total number of time steps.

[0050] A loss function that jointly optimizes the angle error and dynamic smoothness is adopted. Among them, the first term Is used to measure the hip joint angle predicted by the model And the actual hip joint angle The error between them. By minimizing this term, the accuracy of the model's prediction of the hip joint angle can be directly improved. The second term Is used to constrain the change smoothness of the predicted thigh angular velocity . Since in actual motion, the change of the thigh angular velocity should be continuous and smooth, introducing this term can make the trajectory predicted by the model more in line with the physiological characteristics of human motion and avoid unreasonable mutations. As a smoothing factor, it is used to balance the relative importance of the angle error and dynamic smoothness in the loss function. Through a large number of experiments, it is verified that when The value is between 0.1-0.3, it can effectively improve the smoothness of the predicted trajectory while ensuring the prediction accuracy, thereby improving the overall performance of the model.

[0051] To verify the effectiveness of this application, the following experiments were carried out. The publicly available MyoMotion electromyogram signal dataset was used. This dataset contains the electromyogram signals of 50 subjects in 10 different motion states such as walking, running, going up and down stairs, etc., and at the same time records the true values of the corresponding hip joint angles and thigh angular velocities. The deep learning framework uses TensorFlow 2.3.0. In the experiment, the dataset was divided into a training set, a validation set and a test set according to the ratio of 70%, 15%, 15%. The training process of the model uses the Adam optimization algorithm, the learning rate is set to 0.001, and it is trained for 300 rounds. During the training process, an early stopping strategy is used. If the loss value on the validation set does not decrease in 10 consecutive rounds of training, the training is stopped.

[0052] To verify the accuracy of the system's prediction of hip joint movement and the effectiveness of monitoring the thigh movement state, relevant experiments were carried out. As Figure 2Shown are the predicted hip joint angles, actual hip joint angles, and predicted thigh angular velocity curves. Through the experiment, by comparing the predicted hip joint angles, actual hip joint angles, and predicted thigh angular velocity, the performance of the system during the simulation of human lower limb movement is analyzed. This is of great significance for further optimizing the system and improving the rehabilitation effect of the lower limb exoskeleton rehabilitation training robot.

[0053] The process of resolving the thigh-calf co-movement intention based on EMG signals is as follows: Feature extraction and motion intention modeling: The overall steps are the same as those in the process of "resolving the hip-thigh co-movement intention based on EMG signals".

[0054] Among them, the parameter configuration of the weight adjustment factor is determined through experiments as: Knee extension: Quadriceps femoris: = 2.0, because it is the main power source for knee extension, with a contribution degree > 60%. Hamstring muscle (which plays a certain antagonistic regulation role during knee extension): = 0.8, with a contribution degree of about 20%.

[0055] Knee flexion: Hamstring muscle: = 2.2, which is the main power source for knee flexion, with a contribution degree > 65%. Quadriceps femoris (which plays a certain antagonistic regulation role during knee flexion): = 0.7, with a contribution degree of about 20%.

[0056] Hip adduction: Adductor muscles (overall consideration): = 1.8, among which gracilis, pectineus, etc. are the main adductor muscles, with a contribution degree > 55%. Tensor fasciae latae (which has a certain auxiliary adduction effect): = 1.2, with a contribution degree of about 30%.

[0057] Adaptive feature fusion: ; Parameter optimization: = 1.8 (signal-to-noise ratio gain, prefer high-quality signals); = 0.6 (coordination mode gain, strengthening the main contributing muscle groups); Dynamic characteristics: When extending the knee, the weight proportion of the quadriceps femoris > 65%; When flexing the knee, the weight proportion of the hamstring muscle > 58%.

[0058] To visually display the process of adaptive feature fusion, the present invention visualizes the changes in feature weights under different motion states. Taking knee extension as an example, by drawing a three-dimensional chart with time as the horizontal axis, muscle group category as the vertical axis, and weight value as the vertical axis, we get asFigure 3 The three-dimensional graph showing the change of feature weights during knee extension. As Figure 3 shown, at the initial stage of the knee extension movement, the weight value of the quadriceps femoris rapidly rises to about 0.7 and remains at a relatively high level throughout the knee extension process, intuitively reflecting its dominant role in the knee extension movement. The weight value of the hamstring is at a relatively low level in the initial stage of knee extension and gradually rises to about 0.25 as the movement progresses, reflecting the dynamic change of its antagonistic regulation. This visual analysis not only helps to understand the dynamic characteristics of adaptive feature fusion but also provides an intuitive basis for further optimizing the algorithm.

[0059] In order to verify the accuracy of the system's prediction of hip joint movement and the effectiveness of its monitoring of thigh movement state, relevant experiments were carried out. The experiments analyzed the performance of the system during the simulation of human lower limb movement by comparing the predicted hip joint angle, the actual hip joint angle, and the predicted thigh angular velocity. This is of great significance for further optimizing the system and improving the rehabilitation effect of the lower limb exoskeleton rehabilitation training robot.

[0060] The change of knee joint angle is a key index of lower limb movement. Relevant experiments were carried out to obtain the hip joint angle change graph, as specifically shown in Figure 4 the following. Figure 4 It clearly presents the actual hip joint angle and the predicted hip joint angle change curves at different time points, and relatively reliably demonstrates the feasibility of the scheme.

[0061] The process of resolving the co-movement intention of the calf-foot based on EMG signals is as follows: The core algorithm for resolving movement intention: The muscle force distribution matrix: ; The physiological basis of the coefficient matrix is shown in Table 2 below: Table 2 Data table of the physiological basis of the coefficient matrix ; The specific time-varying weight adjustment strategy is as follows: The set formula is as follows: ; Among them, represents the weight coefficient of the i-th row and j-th column in the muscle force distribution matrix at time , which is used to determine the contribution degree of different muscle activation degrees to the output force of the target action and is dynamically adjusted with the change of time and signal quality; is the basic weight, that is, the value in the above table; , representing the adaptive gain coefficient, which is used to determine the sensitivity of the weight coefficient changing with the signal-to-noise ratio; The larger the value, the more drastic the response of the weight coefficient to the change in signal-to-noise ratio; conversely, the smaller the value, the smoother the change in the weight coefficient; is the real-time signal-to-noise ratio of the signal of the i-th channel at a certain moment.

[0062] When the real-time signal-to-noise ratio is greater than 20 dB, the value of is positive, and the weight coefficient will increase on the basis of the basic weight . This means that when the signal quality is good, the contribution weight of each muscle activation degree to the output force of the target action will be appropriately enhanced. For example, if at a certain moment , then , at this time , the weight is increased compared with the basic value.

[0063] When the real-time signal-to-noise ratio is less than 20 dB, [[ID=2,9]] the value of is negative, and the weight coefficient will decrease on the basis of the basic weight . This is because when the signal quality is poor, in order to reduce the influence of noise on the motion intention calculation, the contribution weight of each muscle activation degree is appropriately reduced to avoid misjudgment caused by noise interference. For example, if , then , , the weight is decreased compared with the basic value.

[0064] Through such a time-varying weight adjustment strategy, the weight of the muscle force distribution matrix can be flexibly adjusted under different signal quality conditions, so as to more accurately calculate the motion intention and provide a more reliable input for subsequent impedance control, etc.

[0065] In step 2, the process of the impedance control is specifically as follows: Set the control equation as follows: ; Among them, represents the desired joint angle; represents the desired joint angular velocity; is the actual joint angle; is the actual joint angular velocity; represents the time-varying stiffness; represents the damping coefficient; The proportional term [[ID=,68]] is used to quickly correct the deviation between the desired joint angle [[ID=,71]]and the actual joint angle ; when there is a deviation between the two, the proportional term will generate a torque proportional to the deviation, attempting to reduce this deviation. The larger it is, for the same angular deviation, the greater the corrective torque generated, and the faster the system responds to the angular deviation. However, excessive may cause the system to overshoot and oscillate.

[0066] Differential term , which is used to suppress the oscillation of the system; during the movement, due to the influence of factors such as inertia and friction, the system may exhibit oscillation phenomena. The differential term generates a damping torque proportional to the angular velocity deviation by detecting the desired joint angular velocity and the actual joint angular velocity . When the difference between the actual angular velocity and the desired angular velocity is large, the differential term will generate a large damping force to slow down the movement speed of the joint, thereby suppressing the oscillation. When the knee joint moves too fast when approaching the target angle, it may cause overshooting the target angle and generating oscillation. The differential term will generate a damping torque according to the angular velocity deviation between the two, making the knee joint reach the target angle smoothly.

[0067] Feedforward term , which is used to compensate in advance for the torque required for joint movement according to the muscle force; it calculates the torque that needs to be applied in advance to achieve the desired joint movement under the current movement intention by considering the output force of the muscle force distribution matrix, the tendon force arm and the angle between the muscle pulling line and the joint rotation axis . When performing plantar flexion movement, according to the activation degrees of the lateral gastrocnemius and soleus muscles, the output force is obtained through the muscle force distribution matrix, and then combined with their tendon force arm and angle information, the feedforward term can calculate the corresponding torque in advance, enabling the exoskeleton to better follow the user's movement intention and provide appropriate assistance.

[0068] Among them, the calculation method of the feedforward term is: ; Among them, is the total number of muscle groups participating in the calculation; is the output force of the muscle force distribution matrix; is the tendon force arm; represents the angle between the muscle pulling line and the joint rotation axis.

[0069] The proportional term focuses on quickly correcting the angular deviation, the differential term focuses on suppressing the oscillation during the movement, and the feedforward term compensates in advance for the required torque. The three cooperate with each other to jointly achieve the precise control of the exoskeleton joint movement.

[0070] Such as Figure 5The ankle joint angle change curve is shown as follows. The predicted angle curve and the actual angle curve are highly consistent in terms of the change trend throughout the entire time course. At each time point, the responses of the two to the dorsiflexion or plantar flexion movements of the ankle joint are basically synchronous, indicating that the motion intention recognition method based on EMG signals adopted in the present invention can accurately capture the dynamic trend of ankle joint movement and effectively predict the motion intention of the ankle joint in the time domain.

[0071] For step 3, it is actually a lower limb coordinated control process based on the motion intention calculation result, and the specific process is as follows: 3-1. Motor selection - Brushless DC motor: In the lower limb coordinated control system, a brushless DC motor is selected as the driving element.

[0072] The brushless DC motor has good torque output characteristics. It can provide stable and large torque within a wide speed range. During large-amplitude movements of the hip joint, such as the process from sitting to standing, the motor is required to provide sufficient torque to overcome the body's gravity and joint friction; during the flexion and extension movements of the knee joint and the plantar flexion and dorsiflexion movements of the ankle joint, the motor also needs to adjust the torque output in a timely manner according to different motion states. The brushless DC motor can meet these requirements and ensure the smoothness of joint movement.

[0073] The brushless DC motor has high control precision and can achieve precise speed and position control. Its speed control precision can reach ±1% or even higher, and the position control precision can also meet the strict requirements for joint movement precision in rehabilitation training. During the rehabilitation training process, in order to simulate the walking posture of normal people, it is necessary to accurately control the movement angle and speed of the joint. The high-precision control characteristics of the brushless DC motor can well achieve this goal and help patients carry out more effective rehabilitation training.

[0074] 3-2. Human gait cycle: The walking process of the human body is a complex and orderly periodic movement. Understanding the gait cycle is crucial for fields such as biomechanics research, rehabilitation therapy, and the development of intelligent assistive devices. A complete gait cycle starts from the heel strike of one side and ends when the same-side heel strikes the ground again.

[0075] In the gait cycle, it is mainly divided into the stance phase and the swing phase. The stance phase refers to the stage when the foot is in contact with the ground and bears the body weight, accounting for about 60% of the gait cycle. During this period, it is further subdivided into multiple small stages such as heel strike, flat foot contact, mid-stance, heel off, and toe off, and each small stage has different mechanical characteristics and body posture changes. The swing phase is the stage when the foot leaves the ground and swings forward, accounting for about 40% of the gait cycle, including pre-swing, initial swing, mid-swing, and terminal swing. This stage is mainly for preparing for the next stance.

[0076] Different individuals, exercise states, and health conditions can all affect the gait cycle. For example, the gait cycle of athletes is usually more efficient, while for patients in rehabilitation, due to injuries, the time proportion of their gait cycle and the movements in each stage may be abnormal. Through precise analysis of the gait cycle, it can help doctors evaluate the rehabilitation progress of patients and also assist engineers in designing devices such as exoskeleton robots and intelligent prosthetics that better meet the human movement needs.

[0077] At the same time, relevant experiments were conducted to determine the range of degrees of freedom of each joint, and the angle changes of each joint were obtained as Figure 6 shown. At the same time, according to Figure 6 the results shown, the range of degrees of freedom of each joint is as shown in Table 3 below: Table 3 Data table of the range of degrees of freedom of each joint

[0078] In step 3, the fuzzy adaptive PID control method is as follows: Set an improved fuzzy adaptive PID controller, and the corresponding control law is expressed as: ; where is the output of the controller, used to drive the controlled object; is the angle deviation, corresponding to the difference between the desired angle and the actual angle , that is ; the angle deviation , used to reflect the current control error of the system, is an important basis for the controller to make adjustments; is the deviation change rate, used to reflect the change trend of the error over time; , , are all time-varying control parameters. Among them, corresponds to the proportional, corresponds to the integral, corresponds to the differential coefficient; different from the fixed parameters of the traditional PID controller, these parameters will be adjusted in real time according to the operating state of the system to achieve better control effects.

[0079] The fuzzyfication process uses a triangular membership function to achieve quantization, and the specific formula is as follows: ; where is the center value of the fuzzy subset; is the width of the membership function; is used to convert the angle deviation during the fuzzyfication process Mapped to a fuzzy subset membership function.

[0080] Set the fuzzy rule base to adopt the Mamdani-type rule form with two inputs and three outputs: ; Among them, The k th fuzzy rule in the fuzzy rule base; and are two inputs of the fuzzy controller, is the angle deviation, is the deviation change rate; , , are the three outputs, representing the adjustment amounts of the proportional, integral, and differential coefficients respectively; , , , , are the corresponding fuzzy subsets respectively; Defuzzification is performed by the centroid method, and the specific formula is set as follows: ; Among them, Proportional coefficient adjustment amount; is the number of fuzzy subsets, is , for the th fuzzy subset membership degree; is the representative value of the th fuzzy subset; The centroid method calculates the centroid position of the fuzzy set according to the shape and distribution of the membership function, and takes the centroid position as the precise output after defuzzification; finally, the adjustment amount of the proportional coefficient is obtained; According to the above process, similarly, and can be calculated.

[0081] Online adjustment of the parameters of the PID controller is performed according to the parameter adjustment amounts , and obtained by defuzzification; The parameter adjustment mechanism is as follows: ; Among them, , , respectively represent at time tThe real-time values of the proportional coefficient, integral coefficient, and derivative coefficient of the PID controller at a certain time. As time t changes, these coefficients will be dynamically adjusted according to the operating conditions of the system, so that the controller can better adapt to the characteristics and control requirements of the system, thereby improving the control effect. , and are all the initial values of the PID controller parameters; , , are all adjustment coefficients, which are specifically obtained through experimental calibration; the adjustment coefficients are used to control the amplitude of parameter adjustment and are reasonably set according to the characteristics and control requirements of the system.

[0082] In addition, the present invention also provides a lower limb movement intention recognition system based on electromyogram signals, including; An electromyogram signal acquisition module, which is used to arrange electromyogram sensors on the key hip muscle groups, key knee muscle groups, and key ankle muscle groups of the patient in sequence and collect electromyogram signals; A movement intention prediction module, which is used to perform time-frequency-spatial domain feature fusion and dimensionality reduction processing on the electromyogram signals collected at the key hip muscle groups and key knee muscle groups of the patient, and at the same time use the dynamic weight allocation method and the decoder based on the LSTM-attention mechanism to perform corresponding movement intention prediction, and perform the electromyogram signals collected at the key ankle muscle groups of the patient through the muscle force distribution matrix processing, and at the same time use the time-varying weight adjustment strategy and impedance control to perform corresponding movement intention prediction; A control and regulation module, which is used to control the DC brushless motor on the rehabilitation training device according to the predicted movement intention, and use the fuzzy adaptive PID control method to realize the regulation of joint movement and simulate the walking posture of a real person.

[0083] The present invention performs excellently in movement intention prediction and joint movement control, can effectively overcome the deficiencies of traditional rehabilitation training, improve the safety, comfort and efficiency of rehabilitation training, and provide strong support for the development of lower limb rehabilitation treatment technology.

[0084] The above is only a detailed description of the preferred embodiments and principles of the present invention. For those of ordinary skill in the art, according to the idea provided by the present invention, there will be changes in the specific implementation manners, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. A method for recognizing lower limb movement intention based on electromyogram signals, characterized in that It includes the following steps: S1, arrange sensors at the hip, knee joint and ankle joint of the patient in sequence and collect electromyography signals; S2, perform time-frequency-spatial domain feature fusion and dimensionality reduction processing on the electromyography signals collected at the hip and knee joint parts of the patient, and at the same time use the dynamic weight allocation method and the decoder to perform corresponding motion intention prediction; perform the electromyography signals collected at the ankle joint part of the patient through the muscle force distribution matrix processing, and at the same time adopt the time-varying weight adjustment strategy and impedance control to perform corresponding motion intention prediction; S3, according to the predicted motion intention, the system controls the rehabilitation training equipment and adopts the fuzzy adaptive control method to realize the regulation of joint movement and simulate the walking posture of a real person.

2. The method for identifying lower limb movement intention based on electromyography signals according to claim 1, wherein In step S1, the hip includes: gluteus maximus, gluteus medius / gluteus minimus, and iliopsoas; The knee joint part includes: quadriceps femoris, hamstring muscles, and adductor muscle group; The ankle joint part includes: triceps surae, tibialis anterior, peroneus longus, and peroneus brevis.

3. The lower limb movement intention recognition method based on electromyography signals according to claim 2, wherein In step S2, the time-frequency-spatial domain feature fusion and dimensionality reduction processing includes the following steps: S21, time domain feature extraction: Set the integrated electromyogram value , which is used to quantify the total energy of muscle activation and reflect the intensity of force exertion; set to be the absolute value integral of the electromyogram signal within the time window; Set the root mean square RMS to reflect the instantaneous intensity of dynamic muscle contraction monitoring; S22, frequency domain feature extraction: The median frequency is the frequency point where the power spectrum accumulates 50% energy after Fourier transform of the signal; The power spectral density PSD is calculated by the Welch method in segments to obtain the frequency domain energy distribution, which is used to identify high-frequency noise and low-frequency motion edges; S23, spatial domain feature extraction: Construct a covariance matrix based on the multi-channel electrode array signal, and at the same time identify the deep muscle activation pattern through matrix eigenvalue decomposition; S24, feature fusion and dimensionality reduction: The feature fusion includes the following process: Multi-domain feature splicing: splice the time-domain features , , the main frequency band energy of the frequency-domain features , , and the spatial-domain features into a high-dimensional vector : Normalization processing: Use Z-score standardization to eliminate the dimensional difference of different features; The dimensionality reduction process is as follows: Locally linear embedding: Map the high-dimensional features to the low-dimensional space and retain the topological structure of muscle group co-activation.

4. The method for identifying lower limb movement intention based on EMG signals according to claim 3, wherein In step S2, the dynamic weight allocation method includes the following steps: Muscle group contribution calculation: Dynamically adjust the feature weights according to the task type and output the feature weights of the th muscle group , with a value range of 0-1.

5. The method for recognizing lower limb movement intention based on electromyogram signals according to claim 4, wherein In step S2, the network structure of the decoder is as follows: Input layer, specifically the feature vector after dimensionality reduction ; The LSTM layer, specifically 3 layers of hidden units, is used to capture the temporal dependence of electromyography signals; The attention layer is used to calculate the importance weights of features at each time step; Loss function, specifically a loss function that jointly optimizes the angular error and dynamic smoothness .

6. The method for recognizing lower limb movement intention based on electromyogram signals according to claim 5, characterized in that, In step S2, the muscle force distribution matrix is set as follows: ； : The acting force of the movement of the ankle joint in the plantar direction; : The movement force of the ankle joint in the dorsiflexion direction; : The movement force for the plantar surface to turn inward; : The movement force for the sole of the foot to turn outward.

7. The method for identifying lower limb movement intention based on EMG signals according to claim 6, wherein In step S2, the time-varying weight adjustment strategy is specifically as follows: The set formula is as follows: ； in, Indicates At time t, the weight coefficient of the i-th row and j-th column in the muscle force distribution matrix is used to determine the contribution of different muscle activation levels to the output force of the target movement, and is dynamically adjusted as time and signal quality change; is the basic weight; , represents the adaptive gain coefficient, which is used to determine the weight coefficient Sensitivity to changes in signal-to-noise ratio; The larger the value, the more drastic the response of the weight coefficient to changes in the signal-to-noise ratio; The smaller the value, the smoother the change of the weight coefficient; yes The real-time signal-to-noise ratio of the i-th channel signal at time.

8. The method for identifying lower limb movement intention based on electromyogram signals according to claim 7, characterized in that, In step S2, the process of impedance control is specifically as follows: The set control equation is as follows: ； Among them, represents the desired joint angle; represents the desired joint angular velocity; is the actual joint angle; is the actual joint angular velocity; represents the time-varying stiffness; represents the damping coefficient; Proportional term for quickly correcting the deviation between the desired joint angle and the actual joint angle ; Differential term , used to suppress the oscillation of the system; Feedforward term , which is used to compensate in advance for the torque required for joint movement according to the muscle force.

9. The method for identifying lower limb movement intention based on EMG signals according to claim 8, characterized in that, In step S3, the fuzzy adaptive control method, the specific process is as follows: S31, set an improved fuzzy adaptive PID controller, and the corresponding control law is expressed as: ； Among them, is the output of the controller for driving the controlled object; is the angle deviation, corresponding to the desired angle and the actual angle The difference, that is ; The angle deviation , which is used to reflect the current control error of the system and is an important basis for the controller to make adjustments; is the deviation change rate, which is used to reflect the change trend of the error over time; , , are all time-varying control parameters. Among them, corresponds to the proportion, corresponds to the integral, corresponds to the differential coefficient; S32, the fuzzyfication process is realized by using the triangular membership function for quantization; S33, defuzzification is performed by the centroid method; S34, parameter adjustment amount obtained according to defuzzification 、 and , adjust the parameters of the PID controller online.

10. An electromyogram signal-based lower limb movement intention recognition system for implementing the electromyogram signal-based lower limb movement intention recognition method according to any one of claims 1-9, characterized in that, The lower limb motion intention recognition system based on electromyography signals includes; The electromyography signal acquisition module is used to arrange sensors at the hip, knee joint and ankle joint of the patient in sequence and collect electromyography signals; The motion intention prediction module is used to perform time-frequency-spatial domain feature fusion and dimensionality reduction processing on the electromyography signals collected at the hip and knee joint parts of the patient, and at the same time use the dynamic weight allocation method and the decoder to perform corresponding motion intention prediction, and perform muscle force distribution matrix processing on the electromyography signals collected at the ankle joint part of the patient, and at the same time use the time-varying weight adjustment strategy and impedance control to perform corresponding motion intention prediction; The control and regulation module is used to control the rehabilitation training equipment according to the predicted motion intention, and use the fuzzy adaptive control method to realize the regulation of joint movement and simulate the walking posture of real people.

Images