A wearable motion intent recognition device

Abstract

The application discloses a wearable motion intention recognition device and belongs to the field of motion intention recognition. The wearable device comprises an external material layer, a granular pressure sensor, a flexible contact layer, a master control assembly arranged on the external material layer and a magic tape. The bracelet body is made of two layers of materials, the two layers of materials are insulation layers and can form a ring through the magic tape, so that the bracelet body is conveniently fixed on a limb. A granular pressure sensor array is arranged in the external material layer, and data output by the pressure sensor is preprocessed and then input into a motion intention recognition model for learning. The application can recognize different motion intentions (including direction and size) by detecting the inertia force distribution in the case of small amplitude motion. The application provides a new way for recognizing the motion intention of a patient with weak motion recognition ability and small motion range.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intention recognition, in particular to a wearable motion intention recognition device. BACKGROUND

[0002] Nowadays, with the increasing of the population aging and the hemiplegic patients caused by cerebrovascular diseases, the exoskeleton robot which can provide functions such as motion assistance and rehabilitation training for the wearer has become a research hotspot, but the interaction between the human and the exoskeleton robot requires the robot to actively understand the motion intention of the human. Under the influence of the disease, the patient has very small range of self-controlled limb movement, and the traditional human-computer interaction mode based on program control cannot meet the demand.

[0003] However, the current motion intention recognition method mainly uses multiple fixed cameras to collect multi-directional visual signals. Due to the influence of the disease, the patient's movement ability is weak, and the range of limb movement is small, so it is difficult for the camera to recognize the small movement displacement, which makes the success rate of motion intention recognition low. Moreover, the camera needs an external device to judge the motion intention through visual signals, and the installation position of the camera has specific requirements, which is not suitable for use in an open environment. The visual motion intention recognition method is time-consuming and has low real-time performance. The signal collection process is also affected by obstacles such as shielding, and the device system is complex, and the application environment is limited.

[0004] In summary, the existing motion intention recognition technology mainly faces the following problems: 1) the patient's limb movement range is small, and the signal collection is difficult; 2) the obtained signal has delay and low real-time performance; 3) the signal acquisition device is large and complex, and the use place is limited.

[0005] Therefore, it is necessary to propose a new motion intention recognition device and method to solve the above technical problems. SUMMARY

[0006] In order to overcome the defects of the prior art, the technical problem to be solved by the present application is to provide a motion intention recognition device and method. When the limb moves, the inertia force distribution generated by the wearing of the collection device is collected to obtain the mechanical information of the limb movement direction and size. This wearable device does not need a camera and is easy to wear, portable and convenient.

[0007] To achieve the purpose, the application adopts the following technical solutions: a wearable motion intention recognition device, comprising: an external material layer, a granular pressure sensor array, a flexible contact layer, a master control assembly, and a magic tape; the external material layer is a bendable strip structure, the inside of the strip structure is provided with a groove along the axis, the granular pressure sensor array is arranged in the groove, and then the flexible contact layer is covered on the granular pressure sensor array; the data line of the granular pressure sensor array extends out from the outside of the external material layer, and the master control assembly is arranged at the extending position; one end of the external material layer extends out the magic tape; the motion intention recognition device is bent inward and worn beside the joint position of the human body, and the outside of the external material layer is adhered by the magic tape to realize fastening.

[0008] The processing method in the master control assembly is:

[0009] S1: the tester wears the motion intention recognition device according to the specified position to collect 15 kinds of motion intention data, and transmits the data to the pressure sensor data acquisition and preprocessing module;

[0010] S2: the pressure sensor data acquisition and preprocessing module pre-processes the data to obtain a pressure change data set of gait signals, and extracts pressure signal change characteristics from the data set; meanwhile, the data set is divided into training samples and test samples according to a preset proportion;

[0011] S3: input the obtained data characteristics into the neural network of the self-adaptive enhancement algorithm to train a motion intention recognition model;

[0012] S4: the trained motion intention recognition model is used to recognize the newly collected pressure sensor data.

[0013] Further, the specific steps in step S1 include:

[0014] The granular pressure sensor array is composed of 15 pressure sensors, the tester wears the motion intention recognition device according to the specified position, the wearing position and the wearing method are the same as those in the test, and the specified position is Figure 3The B point on the wearable device coincides with the A point; after the test is completed, the tester performs 5 motion modes most commonly used in daily life at 3 intensities, including: achieving vertical displacement of 0-3 cm, 20-25 cm, and 50-55 cm in a certain motion mode; when the motion intention recognition device is worn near the elbow joint, the 5 motion modes include: elbow flexion, elbow extension, arm lifting, arm lowering, and arm lateral movement; when the motion intention recognition device is worn near the knee joint, the 5 motion modes include: knee flexion, knee extension, leg lifting, leg lowering, and leg lateral movement; a total of 15 intensity and motion mode combinations are obtained, corresponding to 15 motion intentions; wherein the duration of a single intensity and motion mode combination is 1-3 seconds; the pressure data returned by the 5 motion modes in the static state is recorded at the same time;

[0015] Further, the specific steps of data preprocessing in step S2 include:

[0016] Step 2.1: All data at a single sampling time are combined to form a 1X15 original data sample and save the data sample in the initial static state as a comparison signal sequence;

[0017]

[0018] for the kth dimension of the jth original data sample under the ith intensity and motion mode combination, k = 1, 2,..., 15; all original data samples obtained by sampling a single intensity and motion mode combination form a set

[0019] Step 2.2: After filtering the corresponding data in all original data samples, use autocorrelation analysis to find the period m in the original data sample to remove redundant data that repeatedly appears due to the number of winding turns.

[0020] Further, the specific method of step 2.2 is:

[0021] The filtering method is selected as the Kalman filtering method, and the filtered data is subtracted by the pressure data in the static state;

[0022] The de-duplication operation uses autocorrelation analysis to find the period m in the original data sample, and removes the overlapping sensor pressure data in the period, i.e., the data sample changes:

[0023]

[0024] ​Step 2.3: subtract the de-duplicated data sample from the initial static state data sample to obtain the data sample of the dynamic change of pressure; normalize the data sample;

[0025] Further, the specific method of step 2.3 is:

[0026]

[0027] In the formula: is the normalized data of the mthdimensional original data of the jthoriginal sample under the ithintensity exercise mode combination, is the mthdimensional original data of the jthoriginal data sample under the ithgait environment combination, is the maximum value of all mthdimensional original data, is the minimum value of all mthdimensional original data, indicates the mean value of all mthdimensional original data.

[0028] Further, the motion intention recognition model in step S3 includes 105 strong classifiers, each of which corresponds to distinguishing any two of the 15 motion intentions, and the motion intention with the largest number of classification results of the 105 strong classifiers is taken as the output result of the motion intention recognition model; the training method of each strong classifier is:

[0029] Step 3.1: establish a database of basic classifiers, which contains a plurality of basic classifiers;

[0030] Step 3.2: give the same weight to all samples, which represents the probability of selecting the sample for classifier training when training the classifier;

[0031]

[0032] ω t,i is the weight of the ithsample in the classifier at the tthiteration, and N is the number of samples;

[0033] Step 3.3: select samples according to the weight to input each basic classifier, and calculate the classification error rate of each basic classifier, and the basic classifier corresponding to the lowest error rate is a weak classifier;

[0034] Step 3.4: if the samples are correctly classified or exceed the number of iterations, the strong classifier training is completed; otherwise, go to step 3.5;

[0035] Step 3.5: update the weight of the sample, and the updating method is:

[0036] If the current sample is correctly classified:

[0037]

[0038] If the current sample is misclassified:

[0039]

[0040] ω t+1,i is the weight of the i-th sample in the weak classifier in the t+1-th iteration, and t is the error rate of the classifier in the t-th iteration.

[0041] After the update, return to step 3.3.

[0042] Step 3.6: After the training is completed, the classification error rate of all selected weak classifiers is calculated, and the weight of each weak classifier is calculated according to the error rate t The calculation method is:

[0043]

[0044] Step 3.7: The strong classifier includes all weak classifiers and the corresponding weights of each classifier, multiplies the output of each weak classifier by the corresponding weight and accumulates, and judges the classification result according to the accumulation result. The classification result is the output of the strong classifier.

[0045] Compared with the prior art, the beneficial effects of the present application are:

[0046] (1) When the patient's exercise ability is weak and the exercise range is small, the camera cannot accurately identify the small movement, and the delay is strong. The present application itself has a certain weight, and can still generate enough inertial force under small movement, and because the device is worn on the user's limbs, it is real-time and sensitive.

[0047] (2) The wearable device detects the user's movement intention by detecting the inertial force during movement. Because the inertial force is easy to collect and the signal noise is low, the sensor required by the present application is simple to construct and easy to operate, and it is convenient to carry and can be used outdoors.

[0048] (3) The device is connected with magic tape, so that the device can be fixed on the limbs with a certain tension when the user wears it. At the same time, the length of the bracelet is 60 cm, which can be worn on the fingers, arms, thighs, calves and other positions by winding. By overlapping point A on the wearable device and point B on the human body wearing position, the movement intention of these positions can also be detected, so that the portability of the device is strong.

[0049] (4) Multiple devices can be worn on different limbs of the user, and the positions of the devices are determined by the posture sensors built in the devices. For example, on the arm of a human body, the motion intention of the whole arm is obtained by coupling the motion intentions of the upper arm and the lower arm. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The patent subject structure diagram provided in the specific embodiment of the present application is shown in the figure.

[0051] Figure 2 The patent structure diagram provided in the specific embodiment of the present application is shown in the figure.

[0052] Figure 3 The patent installation diagram provided in the specific embodiment of the present application is shown in the figure.

[0053] Figure 4 The human body installation position diagram in the specific embodiment of the present application is shown in the figure.

[0054] Figure 5 The training method diagram of the strong classifier of the present application is shown in the figure.

[0055] In the figure: 1, external material layer; 2, granular pressure sensor; 3, flexible contact layer; 4, master control component device; 5, magic tape. DETAILED DESCRIPTION

[0056] In order to explain the technical content, the purposes and effects of the present application in detail, the following will be described in combination with the embodiments and the accompanying drawings.

[0057] Reference Figures 1-3The application provides a technical scheme: a motion intention recognition device and a recognition method, which comprise an external material layer, a granular pressure sensor array, a flexible contact layer, a master control assembly arranged on the external material layer, and a magic tape. The bracelet body is composed of the external material layer and the flexible contact layer. The inner side of the external material layer has a groove. The inner recess of the external material layer is provided with the granular pressure sensor which is electrically connected with the master control assembly and is arranged in a circular array in the inner recess. In theory, the more pressure sensors are arranged, the more accurate the motion intention detection is. However, considering the size of the pressure sensor and the size of the user's limbs, preferably, fifteen pressure sensors are arranged for detection, so that the recognition accuracy and reliability of the whole device are improved. The lower part is used for accommodating the sensor wire, so that the bracelet volume is smaller and more portable. The flexible contact layer is made of cotton fabric material, which is beneficial to enhancing the use experience of the wearer. When the user's limbs are in motion, the flexible contact layer moves in the normal direction relative to the contacted skin, that is, the flexible contact layer continues to move in the normal direction of the external material layer after being extruded under the action of the inertial force, so that the granular pressure sensor between the external material layer and the flexible contact layer can more directly feel the force during the motion, thereby further preventing information collection failure. After the two layers of materials are fixed together, the magic tape is connected to the other end through the magic tape. The outer side of the external material layer is subjected to sanding treatment and has fine chemical fibers, so that the magic tape can be stuck on the outer side of the material layer at any position, and the tightening and relaxation treatment can be performed to prevent falling. The master control assembly is provided with a posture sensor. The posture sensor senses the current limb position and sends the position of the wearable device to the master controller. This makes the device can detect the position of the limb in real time and has high sensitivity. The length of the external material layer and the flexible contact layer is 50 cm, and the product can be worn on different limbs such as fingers and arms through winding, so that the applicability of the product is wide.

[0058] The inner side groove of the external material layer is provided with a plurality of granular pressure sensors which are arranged in an array. The outer side of the flexible contact layer is connected with the inner side of the external material layer, and the inner side of the flexible contact layer is fixed with a magic tape. The outer side of the external material layer is connected with the master control assembly device and the magic tape. The center of the inner recess outer ring and the master control assembly connection part is provided with a wire outlet and a wire inlet. The outer end of the wire outlet of the external material layer coincides with the outer end of the wire inlet of the master control assembly device:

[0059] When the user wears the motion intention recognition device from the side far from the master component, the user's limbs move to one side, and the flexible contact layer contacts the user's skin. Because the device can be worn on any part of the body limb, the pressure data changes generated by movement in the same direction are similar on the same limb. To make the motion intention recognition model constructed by the collected data applicable to other parts of the limb except the collection site (such as the wrist, etc.), the wearing method at the remaining part of the limb should be consistent with the wearing method at the data collection site, so the application sets a unified wearing method. Point A is located at the upper left corner of the contact surface between the device and the limb, as shown in the schematic Figure 3 illustrated. Point B is located on the side of the limb close to the outside, as shown in the schematic Figure 4 illustrated. The specified position is the overlap of point A of the flexible contact layer and point B on the limb. At the same time, the number of winding turns is selected according to the diameter of the limb, and the magic tape is adjusted to make the wearable device fasten on the user's limb without falling off; the attitude sensor on the master component senses the attitude of the device; at the same time, due to the influence of the weight of the device itself, during the movement, under the action of inertial force, a plurality of granular pressure sensors in contact with the side of the device are extruded, generating a plurality of analog signals which are transmitted to the control component to be converted into digital signals for processing; the pressure sensor data acquisition and preprocessing module sends the collected raw data to the established motion intention recognition model after filtering and other preprocessing.

[0060] The device body is composed of an external material layer and a flexible contact layer, and the length of the external material layer and the flexible contact layer is 50 cm, which can be worn on different limbs such as fingers and arms through winding. And the two layers of materials are fixed together and connected to the other end through a magic tape, and the outside of the external material layer is treated with sanding and has fine chemical fibers. The magic tape can be randomly attached to any position of the external material layer, and the tightness and relaxation process is performed during wearing.

[0061] The further technical features of the application are that the inside of the external material layer has a groove, the upper part of the groove is used to place the granular pressure sensor, the granular pressure sensor is arranged in an array in the recess, and the lower part of the recess is used to accommodate the wiring of the granular pressure sensor.

[0062] The embodiment discloses a motion intention recognition method, which can be used in the wearable device described above, and the motion intention recognition method is implemented according to the following steps:

[0063] S1: let the tester wear the device to collect 15 different data, and transmit the data to the pressure sensor data acquisition and preprocessing module;

[0064] S2: The pressure sensor data acquisition and preprocessing module preprocesses the data to obtain a preprocessed gait signal dataset and extracts features from the data; at the same time, the preprocessed dataset is divided into training samples and test samples according to a preset ratio.

[0065] S3: Input the obtained data features into the neural network of the adaptive enhancement algorithm, train it multiple times to obtain a strong classifier, combine multiple strong classifiers to form a motion intention recognition model, and know the user's limb movement intention based on the accumulator result.

[0066] The specific steps in step S1 include:

[0067] The test subject wore a pressure sensor module consisting of 15 pressure sensors. After wearing the module, the test subject was asked to perform five commonly used movement patterns in daily life at three different intensities. The three intensities included achieving vertical displacement of 0-3cm, 20-25cm, and 50-55cm in a certain movement pattern. The five movement patterns included: elbow flexion, elbow extension, arm raising, arm lowering, and arm lateral movement, resulting in a total of 15 combinations of intensities and movement patterns. The duration of each combination of intensity and movement pattern was 1-3 seconds. At the same time, the pressure data returned by the five movement patterns were recorded when the test subject was at rest.

[0068] The specific steps of data preprocessing in step S2 include:

[0069] All data from a single sampling moment constitute a 1x15 original data sample. For the j-th original data sample under the i-th combination of intensity and motion mode The raw data in the k-th dimension, k = 1, 2, ... 15. The 15-dimensional data are arranged in a clockwise direction, starting from the sensor at point A, with the arm wrapped around it. All raw data samples obtained from combined sampling of single intensity and motion patterns are also included. The set formed is

[0070] Each data acquisition and preprocessing module handles x raw All corresponding data in the original data samples undergo filtering, deduplication, and normalization preprocessing. The Kalman filter method is selected for filtering. After filtering, the pressure data under static conditions is subtracted. Deduplication is performed using autocorrelation analysis to find the period *m* within the original data samples. Overlapping sensor pressure data is then removed based on this period. (Single original data sample) The normalization method for each dimension of the data is as follows:

[0071]

[0072] In the formula: is the normalized data of the mth-dimensional original data of the jth original sample under the ith intensity exercise mode combination, is the mth-dimensional original data of the jth original data sample under the ith gait environment combination,

[0073] is the maximum value of all the mth-dimensional original data, is the minimum value of all the mth-dimensional original data, indicates the mean value of all the mth-dimensional original data; then the normalized data is used to obtain time sequence features (such as mean value, variance, time domain statistical features), frequency domain features (such as Fourier transform) and the like.

[0074] The specific steps of obtaining a strong recognition model through multiple training in the step S3 include:

[0075] In the exercise intention recognition model, there are 15 categories of sample classification, which are respectively the first intensity of elbow flexion, the second intensity of elbow flexion,..., and they are named as 1, 2, 3,..., 15. Each category is provided with an accumulator, and the accumulator is incremented by one when the sample is determined to be this category. The exercise intention recognition model mainly uses the adaptive boosting binary classification algorithm, and 1 is divided into the first category and set as label a, and 2 is selected as the second category from the remaining 14 categories and set as label b. The total number of samples is N, and the training samples are assigned with the same weight In the input multi-scale convolutional neural network based on the adaptive boosting algorithm, a first weak classifier and a classification result are obtained through multiple sample training. In the specific training process, if a sample point has been accurately classified, its weight is reduced when the next training set is constructed; on the contrary, if a sample point has not been accurately classified, its weight is increased. Then, the sample set with updated weight is used to train the next classifier, and the above steps are repeated to obtain a plurality of weak classifiers. Finally, for the weak classifiers with low misclassification rate, higher weights are given, and the plurality of weak classifiers are combined into a strong classifier.

[0076] The specific steps of combining a plurality of strong classifiers to constitute the exercise intention recognition model in the step S3 include:

[0077] The 15 categories of test samples are traversed, 1 is taken as the first category, 3 is selected as the second category, the above process is repeated, and a strong classifier is obtained again. 1 and 4, 1 and 5,..., are taken again. Finally, 105 strong classifiers can be obtained. A test sample is input into the 105 strong classifiers for learning, and when the process ends, the accumulators of the 15 categories are checked to determine that the test sample belongs to the category with the maximum accumulator value.

[0078] ​The application is described by preferred embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to the features and embodiments without departing from the spirit and scope of the application. The application is not limited by the specific embodiments disclosed herein, and other embodiments falling within the claims of the application are within the scope of the application.

Claims

1. A wearable motion intent recognition device, the recognition device comprising: The external material layer is a flexible strip structure, the inside of the strip structure is provided with a groove along the axis, the granular pressure sensor array is arranged in the groove, and then the flexible contact layer is covered on the granular pressure sensor array; the data line of the granular pressure sensor array extends out of the outside of the external material layer, and the main control assembly is arranged at the extending position; one end of the external material layer extends out of the magic tape; the motion intention recognition device is bent inward and worn beside the joint position of the human body, and the outside of the external material layer is adhered through the magic tape, so that the wearing is fastened; The processing method in the main control assembly is: S1: let the tester wear the motion intention recognition device according to the specified position to collect 15 kinds of motion intention data, and transmit the data to the pressure sensor data acquisition and pretreatment module; The granular pressure sensor array is composed of 15 pressure sensors, and the tester wears the motion intention recognition device according to the specified position, and the wearing position and wearing method are the same as those during the test; After the test is completed, the tester performs 5 kinds of motion modes most commonly used in daily life at 3 intensities, the 3 intensities include: achieving vertical displacement of 0-3cm, 20-25cm and 50-55cm in a certain motion mode; when the motion intention recognition device is worn near the elbow joint, the 5 kinds of motion modes include: elbow flexion, elbow extension, arm lifting, arm lowering and arm side; when the motion intention recognition device is worn near the knee joint, the 5 kinds of motion modes include: knee flexion, knee extension, leg lifting, leg lowering and leg side; 15 intensity and motion mode combinations are obtained, corresponding to 15 kinds of motion intentions; wherein the duration of a single intensity and motion mode combination is 1-3 seconds; the pressure data returned by the 5 kinds of motion modes in the static state is recorded at the same time; S2: the pressure sensor data acquisition and pretreatment module pretreats the data to obtain a pressure change data set of the motion signal, and extracts the pressure signal change characteristics from the data set; meanwhile, the data set is divided into training samples and test samples according to a preset proportion; The specific steps of data pretreatment in step S2 include: Step 2.1: All data of a single sampling instant is grouped into a 1X15 raw data sample and the data sample of the initial rest state is saved as a comparison signal sequence; the kth dimension of the jth raw data sample under the ith intensity and motion pattern combination, k = 1, 2,... 15; all raw data samples obtained from a single intensity and motion pattern combination sample ​​ Step 2.2: after filtering all corresponding data in the original data samples, autocorrelation analysis is used to find the period m in the original data samples, and redundant data repeatedly appearing due to winding turns is removed; Step 2.3: subtract the data sample after removing the redundancy from the data sample in the initial static state to obtain the data sample in the pressure dynamic change; the data sample is normalized; S3: input the obtained data characteristics into the neural network of the self-adaptive enhancement algorithm to train a motion intention recognition model; S4: use the trained motion intention recognition model to recognize the newly collected pressure sensor data.

2. The wearable motion intent recognition device of claim 1, wherein, The motion intention recognition model in the step S3 includes 105 strong classifiers, each corresponding to distinguishing any two of the 15 motion intentions, and the motion intention with the largest number of classification results of the 105 strong classifiers is taken as the output result of the motion intention recognition model; the training method of each strong classifier is: Step 3.1: a database of basic classifiers is established, and the database includes a plurality of basic classifiers; Step 3.2: the same weight is given to all samples, and the weight represents the probability of selecting the sample for classifier training during classifier training; ω t,i ωt is the weight of the ith sample in the classifier at the tth iteration, and N is the number of samples. Step 3.3: the samples are input into each basic classifier according to the weight, and the classification error rate of each basic classifier is calculated, and the basic classifier corresponding to the lowest error rate is a weak classifier; Step 3.4: if the samples are correctly classified or the iteration number exceeds, the strong classifier training is ended; otherwise, step 3.5 is entered; Step 3.5: the weight of the sample is updated, and the updating method is: If the current sample is correctly classified: If the current sample is incorrectly classified: ω t+1,i is the weight of the i-th sample in the weak classifier for the t+1 iteration, ε t is the error rate of the classifier for the t iteration; After the updating is completed, step 3.3 is returned; Step 3.6: Calculate the classification error rate of all selected weak classifiers after the training, and calculate the weight α of each weak classifier according to the error rate t The calculation method is: Step 3.7: the strong classifier includes all weak classifiers and the weights corresponding to the classifiers, the outputs of the weak classifiers are multiplied by the corresponding weights and accumulated, and the classification result is judged according to the accumulated result, and the classification result is the output of the strong classifier.

3. The wearable motion intent recognition device of claim 1, wherein, The specific method of the step 2.2 is: The filtering method is selected as the Kalman filtering method; The de-duplication operation is using autocorrelation analysis to find the period m within the original data sample, removing the overlapping part of the sensor pressure data, i.e. the data sample Changes occur:

4. The wearable motion intent recognition device of claim 1, wherein, The specific method of the step 2.3 is: wherein: is the normalized data of the mth dimension of the jth original data sample under the ith intensity exercise mode combination, is the mth dimension of the jth original data sample under the ith intensity exercise mode combination, is the maximum value of all the mth dimension of the original data, is the minimum value of all the mth dimension of the original data, represents the mean value of all the mth dimension of the original data.

Images