1. A motion state identification method of an upper limb assistance exoskeleton robot is characterized by comprising the following steps:

step 1: classifying the finite motion states of the exoskeleton of the robot according to a Finite State Machine (FSM), and making a finite motion state conversion relation;

the robot exoskeleton limited motion states are divided into three types: a rotation assisting state which helps the operator to rotate the shoulder and the arm to exert force; the power-assisted state is maintained, which helps the operator to keep exerting force through the shoulder and arm fixing action; relaxed state-corresponding to a relaxed state in which the operator does not perform any heavy physical work;

the robot exoskeleton can be in one state at any moment, the rotation assistance state and the assistance maintaining state can be mutually reached, the rotation assistance state and the relaxation state can be mutually reached, and the assistance maintaining state and the relaxation state can not be mutually reached;

step 2: installing a plurality of sensors on the robot, and acquiring angular velocity signals, angular acceleration signals and interaction force signals;

data signals collected from sensorsPreprocessing is carried out to obtain a data state w_jI is the serial number of the sensor, and j is the time; according to the data state w_jClassifying the limited motion state of the exoskeleton of the robot at the moment j by adopting an HSVM algorithm;

and step 3: utilizing majority of data states within a sliding window of size l along a time axis of sensor acquisition robot dataThe voter mechanism takes the category corresponding to the most number of the same data states in a sliding window as the estimated value k of the current system state_m，k_mTaking a value as one of the limited motion states of the exoskeleton of the three types of robots, wherein m is a time window serial number;

and 4, step 4: judging whether the front and back states of the system are reachable, if so, judging whether the front state k of the system is reachable_m-1Estimate k to the current state_mThe reachable state, the system state ss at this time_mFrom k to k_m-1To k_mAnd returning to the step 2, otherwise, executing the step 5;

and 5: the HSVM algorithm is adopted to carry out category judgment again on the same data state with the largest quantity in the current sliding window, and a new system state estimated value k 'is generated'_mAnd again check k'_mWhether or not it is k_m-1If yes, returning to step 2, otherwise, reporting the system state as an abnormal value ss_mRepeating the step 5 and counting the times of the abnormal state; and when the occurrence frequency of the abnormal state exceeds a set threshold value, stopping the robot from working.

2. The method of claim 1, wherein the sensors mounted on the robot comprise an inertial measurement unit and a membrane pressure sensor.

3. The method as claimed in claim 2, wherein the inertial measurement unit is installed at left and right brachial biceps positions, the included angle of the upper arm with respect to the vertical direction is the rotation angle of the upper arm, and when the upper arm is located at the front side of the vertical axis body, the angle is positive, otherwise, the angle is negative.

4. The method of claim 2, wherein the membrane pressure sensors are installed at the left and right brachial biceps and brachial triceps longhead positions.

5. The method for identifying the motion state of the upper limb assistance exoskeleton robot according to claim 2, wherein the model of the inertial measurement unit is an MPU 6050.

6. The method for recognizing the motion state of the upper limb assistance exoskeleton robot according to claim 1, wherein the preprocessing the data signals collected by the sensors comprises:

for the interaction force signal, denoising by adopting a wavelet transform method based on a threshold value;

performing Kalman filtering denoising on the angular velocity signal and the angular acceleration signal, and performing fusion solution on the data to calculate a large-arm rotation angle value; and extracting features by adopting a wavelet packet transformation-based and statistic-based combined feature extraction method.

7. The method for recognizing the motion state of the upper limb assistance exoskeleton robot according to claim 1, wherein the HSVM algorithm is a one-to-one support vector machine ovo-SVM optimized by a harmony search algorithm HS, and the method comprises the following specific steps:

step 2-1: the main control parameters of the initialization harmony search algorithm HS are:

the capacity of the harmony memory library HM, namely the number HMS of the initial harmony, and the value of the harmony memory library HM is related to the searching capability and the optimization performance of the algorithm;

the harmony memory bank reserves the probability HMCR, namely the probability of selecting solution components from the harmony memory bank when a new solution is generated, and the value relationship of the probability is the diversity of harmony and the acquisition of the optimal solution;

the memory bank disturbance probability PAR is the probability of fine tuning disturbance to the partial components each time;

local disturbance broadband u, the value of which influences the performance and solving precision of the algorithm;

step 2-2: setting a target function, and selecting the classification accuracy when training the model as the target function:

f(x)＝V_acc (1)

wherein, V_accThe classification accuracy is;

step 2-3: setting key parameters to be optimized and a search range, namely an oo-SVM penalty factor C and a kernel function parameter sigma of a one-to-one support vector machine optimized by a harmony search algorithm HS;

step 2-4: robot data acquired by a sensor is input into an ovo-SVM algorithm, and parameters of an iterative loop of the robot data are optimized by using an HS algorithm.

Background

The upper limb assistance exoskeleton robot has already made a certain progress in the aspects of mechanical structure and motion control algorithm, but how to effectively arrange a sensor network of the exoskeleton robot, how to acquire motion information of an exoskeleton in real time, and how to accurately judge the motion state of the exoskeleton in real time, so that the assistance process is ensured to be performed smoothly, stably and safely, and the upper limb assistance exoskeleton robot is still a key point and a difficulty point in the design process of the upper limb assistance exoskeleton robot.

At present, in the aspect of an exoskeleton motion state identification method, a certain research is carried out on the exoskeleton by the university of Zhejiang through pottery and jungle (pottery, pneumatic power-assisted exoskeleton robot and robot coordinated motion control research [ D ]. university of Zhejiang, 2018.), an upper limb part of a power-assisted exoskeleton adopts myoelectric signals and a support vector machine to carry out motion state identification and judgment, but human myoelectric signals are weak, a myoelectric signal sensor is easily interfered by external factors, and the signal precision is not high; in addition, the key parameters of the support vector machine are set in advance, so that the optimal parameters cannot be selected according to various actions in the boosting process, and the adaptability is low. The invention patent CN112336340A discloses a human motion intention identification method for a waist assistance exoskeleton robot, which can accurately distinguish four motion states of bending, standing, left leg walking and right leg walking, but does not consider whether the conversion between the motion states has safety problems or not.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides a motion state identification method of an upper limb assistance exoskeleton robot, which combines a pair of support vector machines (ovo-SVM) optimized by Harmonic Search (HS) with a Finite State Machine (FSM), firstly classifies the finite motion state of the exoskeleton of the robot according to the FSM of the finite state machine, and formulates a finite motion state conversion relation; secondly, collecting robot data through a sensor and classifying the states of the robots, and using a majority voter mechanism for the data states in the sliding window along a time axis to take the most data states as the estimated values of the current system state; and then judging the accessibility of the front and rear states of the system, and judging the category of the same data state with the largest quantity in the current sliding window again by adopting an HSVM algorithm to determine the final motion state of the robot. The method of the invention can not only improve the classification accuracy, but also change the parameter value according to the change of the condition, and has stronger robustness and good self-adaptability to the uncertain factors of the system.

The technical scheme adopted by the invention for solving the technical problem comprises the following steps:

step 1: classifying the finite motion states of the exoskeleton of the robot according to a Finite State Machine (FSM), and making a finite motion state conversion relation;

the robot exoskeleton limited motion states are divided into three types: a rotation assisting state which helps the operator to rotate the shoulder and the arm to exert force; the power-assisted state is maintained, which helps the operator to keep exerting force through the shoulder and arm fixing action; relaxed state-corresponding to a relaxed state in which the operator does not perform any heavy physical work;

the robot exoskeleton can be in one state at any moment, the rotation assistance state and the assistance maintaining state can be mutually reached, the rotation assistance state and the relaxation state can be mutually reached, and the assistance maintaining state and the relaxation state can not be mutually reached;

step 2: installing a plurality of sensors on the robot, and acquiring angular velocity signals, angular acceleration signals and interaction force signals;

data signals collected from sensorsPreprocessing is carried out to obtain a data state w_jI is the serial number of the sensor, and j is the time; according to the data state w_jClassifying the limited motion state of the exoskeleton of the robot at the moment j by adopting an HSVM algorithm;

and step 3: utilizing a majority voter mechanism for data states in a sliding window with the size of l along a time axis of collecting robot data by a sensor, and taking the category corresponding to the same data state with the largest quantity in the sliding window as a current system state estimated value k_m，k_mTaking a value as one of the limited motion states of the exoskeleton of the three types of robots, wherein m is a time window serial number;

and 4, step 4: judging whether the front and back states of the system are reachable, if so, judging whether the front state k of the system is reachable_m-1Estimate k to the current state_mThe reachable state, the system state ss at this time_mFrom k to k_m-1To k_mAnd returning to the step 2, otherwise, executing the step 5;

and 5: the HSVM algorithm is adopted to carry out category judgment again on the same data state with the largest quantity in the current sliding window, and a new system state estimated value k 'is generated'_mAnd again check k'_mWhether or not it is k_m-1If yes, returning to step 2, otherwise, reporting the system state as an abnormal value ss_mRepeating the step 5 and counting the number of times of abnormal states; and when the occurrence frequency of the abnormal state exceeds a set threshold value, stopping the robot from working.

Further, the robot-mounted sensor includes an inertial measurement unit and a membrane pressure sensor.

Further, the inertia measurement unit is installed at the positions of the biceps brachii muscle of the left and right large arms, the included angle of the large arms relative to the vertical direction is a large arm rotation angle, when the large arms are positioned on the front side of the vertical axis body, the angle is a positive value, and otherwise, the angle is a negative value.

Further, the membrane pressure sensors are installed at the positions of the left and right brachial biceps brachii and the long head of the brachial triceps.

Further, the inertial measurement unit is an MPU 6050.

Further, the preprocessing the data signal collected by the sensor comprises:

for the interaction force signal, denoising by adopting a wavelet transform method based on a threshold value;

performing Kalman filtering denoising on the angular velocity signal and the angular acceleration signal, and performing fusion solution on the data to calculate a large-arm rotation angle value; and extracting features by adopting a wavelet packet transformation-based and statistic-based combined feature extraction method.

Further, the HSVM algorithm is a one-to-one support vector machine ovo-SVM optimized for the harmony search algorithm HS, and the specific steps are as follows:

step 2-1: the main control parameters of the initialization harmony search algorithm HS are:

the capacity of the harmony memory library HM, namely the number HMS of the initial harmony, and the value of the harmony memory library HM is related to the searching capability and the optimization performance of the algorithm;

the harmony memory bank reserves the probability HMCR, namely the probability of selecting solution components from the harmony memory bank when a new solution is generated, and the value relationship of the probability is the diversity of harmony and the acquisition of the optimal solution;

the memory bank disturbance probability PAR is the probability of fine tuning disturbance to the partial components each time;

local disturbance broadband u, the value of which influences the performance and solving precision of the algorithm;

step 2-2: setting a target function, and selecting the classification accuracy when training the model as the target function:

f(x)＝V_acc (1)

wherein, V_accThe classification accuracy is;

step 2-3: setting key parameters to be optimized and a search range, namely an oo-SVM penalty factor C and a kernel function parameter sigma of a one-to-one support vector machine optimized by a harmony search algorithm HS;

step 2-4: robot data acquired by a sensor is input into an ovo-SVM algorithm, and parameters of an iterative loop of the robot data are optimized by using an HS algorithm.

The invention has the following beneficial effects:

(1) the method of the invention can not only improve the classification accuracy, but also change the parameter value according to the change of the condition, and has stronger robustness and good self-adaptability to the uncertain factors of the system.

(2) The invention introduces a Finite State Machine (FSM) concept into an upper limb assistance exoskeleton robot, and provides a motion state identification method (F-HSVM) which is provided by combining a pair of support vector machines (ovo-SVM) optimized by Harmonic Search (HS) and the FSM. The finite state machine can optimize the classification effect of the basic classifier, the motion states of the exoskeleton are divided into a limited number, and the system can only be in a certain state at any time. When a specific event or condition is triggered, state transition is carried out between the reachable states; the state change which is not in the limited state is regarded as an abnormal state, the state conversion is not carried out, and the exoskeleton directly enters a stop state to protect an operator, so that the safety of the exoskeleton assistance is improved.

Drawings

FIG. 1 is a flow chart of the F-HSVM method of the present invention.

Fig. 2 is a transition diagram of the finite state machine of the upper limb assistance exoskeleton robot.

FIG. 3 is a diagram illustrating the classification accuracy of the F-HSVM method of the present invention.

FIG. 4 is a comparison diagram of classification accuracy of the ovo-SVM algorithm of the present invention.

FIG. 5 is a diagram illustrating the detection of abnormal values of F-HSVM according to the present invention.

Detailed Description

The invention is further illustrated with reference to the following figures and examples.

The invention aims to overcome the defects in the prior art and provides the motion state identification method of the upper limb assistance exoskeleton robot, which is high in man-machine interaction level, accurate in motion state identification effect and high in safety coefficient.

As shown in fig. 1, a method (F-HSVM) for recognizing a motion state of an upper limb assistance exoskeleton robot includes the following steps:

step 1: classifying the finite motion states of the exoskeleton of the robot according to a Finite State Machine (FSM), and making a finite motion state conversion relation;

as shown in fig. 2, the limited motion states of the robot exoskeleton can be divided into three categories: a rotation assisting state which helps the operator to rotate the shoulder and the arm to exert force; the power-assisted state is maintained, which helps the operator to keep exerting force through the shoulder and arm fixing action; relaxed state-corresponding to a relaxed state in which the operator does not perform any heavy physical work;

the robot exoskeleton can be in one state at any moment, the rotation assistance state and the assistance maintaining state can be mutually reached, the rotation assistance state and the relaxation state can be mutually reached, and the assistance maintaining state and the relaxation state can not be mutually reached;

step 2: installing a plurality of sensors on the robot, and acquiring angular velocity signals, angular acceleration signals and interaction force signals;

data signals collected from sensorsPreprocessing is carried out to obtain a data state w_jI is the serial number of the sensor, and j is the time; according to the data state w_jClassifying the limited motion state of the exoskeleton of the robot at the moment j by adopting an HSVM algorithm;

and step 3: utilizing a majority voter mechanism for data states in a sliding window with the size of l along a time axis of collecting robot data by a sensor, and taking the category corresponding to the same data state with the largest quantity in the sliding window as a current system state estimated value k_m，k_mTaking a value as one of the limited motion states of the exoskeleton of the three types of robots, wherein m is a time window serial number;

and 4, step 4: judging whether the front and back states of the system are reachable, if so, judging whether the front state k of the system is reachable_m-1Estimate k to the current state_mThe reachable state, the system state ss at this time_mFrom k to k_m-1To k_mAnd returning to the step 2, otherwise, executing the step 5;

and 5: the HSVM algorithm is adopted to carry out category judgment again on the same data state with the largest quantity in the current sliding window, and a new system state estimated value k 'is generated'_mAnd again check k'_mWhether or not it is k_m-1If yes, returning to step 2, otherwise, reporting the system state as an abnormal value ss_mRepeating the step 5 and counting the number of times of abnormal states; and when the occurrence frequency of the abnormal state exceeds a set threshold value, stopping the robot from working.

Further, the robot-mounted sensor includes an inertial measurement unit and a membrane pressure sensor.

Further, the inertia measurement unit is installed at the positions of the biceps brachii muscle of the left and right large arms, the included angle of the large arms relative to the vertical direction is a large arm rotation angle, when the large arms are positioned on the front side of the vertical axis body, the angle is a positive value, and otherwise, the angle is a negative value.

Further, the membrane pressure sensors are installed at the positions of the left and right brachial biceps brachii and the long head of the brachial triceps.

Further, the inertial measurement unit is an MPU 6050.

Further, the preprocessing the data signal collected by the sensor comprises:

for the interaction force signal, denoising by adopting a wavelet transform method based on a threshold value;

performing Kalman filtering denoising on the angular velocity signal and the angular acceleration signal, and performing fusion solution on the data to calculate a large-arm rotation angle value; and extracting features by adopting a wavelet packet transformation-based and statistic-based combined feature extraction method.

Further, the HSVM algorithm is a one-to-one support vector machine ovo-SVM optimized for the harmony search algorithm HS, and the specific steps are as follows:

step 2-1: the main control parameters of the initialization harmony search algorithm HS are:

the capacity of the harmony memory library HM, namely the number HMS of the initial harmony, and the value of the harmony memory library HM is related to the searching capability and the optimization performance of the algorithm;

the harmony memory bank reserves the probability HMCR, namely the probability of selecting solution components from the harmony memory bank when a new solution is generated, and the value relationship of the probability is the diversity of harmony and the acquisition of the optimal solution;

the memory bank disturbance probability PAR is the probability of fine tuning disturbance to the partial components each time;

local disturbance broadband u, the value of which influences the performance and solving precision of the algorithm;

step 2-2: setting a target function, and selecting the classification accuracy when training the model as the target function:

f(x)＝V_acc (1)

wherein, V_accThe classification accuracy is;

step 2-3: setting key parameters to be optimized and a search range, namely an oo-SVM penalty factor C and a kernel function parameter sigma of a one-to-one support vector machine optimized by a harmony search algorithm HS;

step 2-4: robot data acquired by a sensor is input into an ovo-SVM algorithm, and parameters of an iterative loop of the robot data are optimized by using an HS algorithm.

The specific embodiment is as follows:

example 1:

as shown in FIG. 3, the F-HSVM method is subjected to experimental simulation analysis of classification accuracy by using MATLAB/Simulink, and a control experiment is performed by using an ov-SVM algorithm for effective comparison.

In this example, the data is coherent motion data, and the motion state distribution sequence is 0 → 1 → 2 → 1 → 2 → 1 → 0(0 represents the relaxed state, 1 represents the rotation assist state, and 2 represents the retention assist state). In fig. 4 and 5, the abscissa represents the sampling points, the ordinate represents the classification label, "·" represents each sampling point and its corresponding actual state value, and ". smallcircle" represents each sampling point and its corresponding predicted state value. The comparison shows that the identification rate of the ovo-SVM algorithm to the relaxed state and the rotation assistance state is low, the phenomenon that the rotation assistance is sometimes generated in practical application is easy to occur, the F-HSVM method can accurately classify most motion states, the classification accuracy is high, and the identification effect is stable. Through verification, the classification accuracy of the F-HSVM is 99.212%, and the classification accuracy after the over-SVM replacement is 85.531%.

Example 2

And (3) checking the processing capacity of the F-HSVM method for abnormal values by using MATLAB/Simulink, and performing a simulation experiment on the data set added with abnormal state conversion.

In this example, the data is the exercise data added with abnormal state conversion, i.e. the data of the relaxed state and the assistance-maintaining state are communicated. The state distribution sequence is 0 → 1 → 2 → 0 → 2 → 1 → 0, where 0 (relaxed state) and 2 (hold boost state) are forbidden to reach in the finite state machine, in which case the algorithm should turn the system state to 3 (warning state) and the system is stopped urgently when the number of consecutive warning states reaches a threshold (set to 20 in the experiment).

As can be seen from FIG. 5, the algorithm can predict the motion state more accurately for the sampling points of about 0-150. When the state of the provided sample point changes from 2 to 0, the transition is prohibited, so the algorithm classifies it as 3, i.e., a warning state. In fig. 5, the system automatically enters a stop state when the warning state reaches 20 consecutive times, and the program stops operating. Therefore, the algorithm can effectively identify the situation of the inaccessible motion state under the condition of ensuring the classification accuracy, and the safety of the system is improved.