Flexibly placed textile sensors for robust joint motion tracking

Abstract

The present invention discloses a textile sensor for robust joint motion tracking that can be flexibly placed. The textile sensor is composed of a standard textile elbow pad and a sparse network of six flexible stretch sensors, including a preparation process of the textile sensor: winding silk fibers around yarn. The present invention is directed to existing methods that mainly classify specific movements rather than accurately reconstructing human body movements using flexible sensors, and generally cannot compensate for large displacements. Some methods do not support sensor rotation. Due to the deformable characteristics of flexible sensors, it is more difficult to achieve robust motion tracking when the sensor has a position offset. In contrast, the method of the present invention can not only robustly track human joint movements, but also be robust to large annular and lateral device displacements. In addition, since the predictor of the present invention can compensate for displacements through a robust algorithm, the method of the present invention does not require calibration.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to flexibly placed textile sensors for robust joint motion tracking. Background Art

[0002] Human motion tracking has great potential in understanding human intentions and actively controlling physical devices and systems, and has been actively explored by the public. As a new type of human-computer interaction, human motion tracking can be applied to many fields such as robotics, haptics, biomechanical research, rehabilitation, healthcare, and entertainment. Current motion capture systems based on optical tracking or inertial sensors are widely used in the film, animation, and gaming industries. However, optical solutions have poor capture performance under severe occlusion, poor lighting conditions, or outdoor environments. Motion capture systems based on inertial measurement units (IMUs) can be freely used in outdoor scenes, but their tracking accuracy is very sensitive to the placement of their sensors. Such IMU sensors must be tightly fixed to the human body with straps to ensure stable tracking. However, this can easily lead to an uncomfortable experience for users. In addition, before using IMU sensors for tracking, users must calibrate them, and if one sensor deviates from its initial position, this calibration process must be repeated. Therefore, there is still a lack of a stable method to accurately capture human motion that supports flexible placement of sensors.

[0003] The proposed smart elbow pad is able to accurately and robustly track human motion despite considerable variations in the placement of these tracking devices. The elbow joint was chosen as the measurement joint for the prototype because athletes often use elbow pads to protect them from injuries when they fall or are hit. The learning-based approach of this work can also be effectively applied to other joints by appropriately modifying the equipment and procedures of this work. The method eliminates the strict constraints of precise placement of the device on the body and allows the user to freely wear the prototype in different usage scenarios like wearing a normal elbow pad. The flexible device placement has a great impact on the sensor readings and poses significant challenges to the corresponding sensor signal processing algorithms. The core contribution of this work is to achieve a low-cost, robust and stable solution for human joint motion tracking with different device displacements. To this end, this paper proposes a textile sensor for robust joint motion tracking that can be flexibly placed. Summary of the invention

[0004] It is an object of the present invention to provide a flexibly positionable textile sensor for robust joint motion tracking to address the above-mentioned deficiencies in the art.

[0005] In order to achieve the above object, the present invention provides the following technical solution: the textile sensor is composed of a standard textile elbow guard and a sparse network of six flexible stretch sensors, which includes the preparation process of the textile sensor:

[0006] S1: winding silk fibers into yarn;

[0007] S2: Based on S1, the polyurethane solution is diluted to 6.5 mg / ml, and mixed with 5 mg / ml silver nanowire solution in equal volume ratio, and finally, a 5.2 mg / ml nanosilver wire solution is poured on the surface of the woven yarn structure by the dipping-drying method, and then poured on the surface after dipping and drying;

[0008] S3: Based on S2, the head of the silver thread (cotton fiber impregnated with silver ions) is fixed to the end of the sensor by sewing;

[0009] S4: Based on S3, wrap the end of the silver wire around the head of the conductor and wrap it with insulating glue;

[0010] S5: Based on S4, the end of the conductor is directly soldered to the circuit board and sealed with hot melt adhesive. When the sensor is stretched, the conductive fibers separate and the contact area decreases, resulting in an increase in resistance. The sensor is then hand-sewn into the elbow pad using a flat needle and connected to the circuit board via a conductive silver wire (cotton fiber impregnated with silver ions).

[0011] As a preferred solution of the present invention, the textile sensor also includes a process of collecting data from the sensor using a six-channel PCB board:

[0012] Data is sent at a speed of 20 frames per second (FPS) through the Bluetooth chip (Nrf51822). The channels are selected in sequence by the multiplexer voltage divider. The difference between the voltage of each sensor and the reference voltage is calculated using the Wheatstone bridge structure. Vref = Vcc / 2, where Vcc represents the voltage of the access circuit. A pair of identical resistor units (represented by the symbol R) is also used to obtain the reference voltage. The input voltage V of the analog-to-digital conversion adc c cin is defined as (ignoring the effect of the low-pass filter):

[0013]

[0014] Among them, R i Represents the voltage divider resistor, Rsensor i Indicates the i th A flexible sensor, Gain represents the amplification factor of the amplifier;

[0015] The voltage measurement results are processed by a low-pass filter with a bandwidth of 300Hz, converted into digital form within [0, 1023], and transmitted through the Bluetooth antenna. The designed circuit board sends the digitized sensor signal to the mobile terminal via Bluetooth, and then the data is further transmitted to the remote server, which executes the calculation process of the LightGBM model. The circuit can also send data directly to the server via a Bluetooth adapter. The server is configured with a single-core CPU (no GPU), 2GB of memory, and the operating system is Ubuntu16.04 64-bit.

[0016] As a preferred solution of the present invention, the textile sensor further includes a data set acquisition process:

[0017] The dataset preparation process is divided into three stages:

[0018] ① A single user performs activities at different device offset positions, and the collected data D ss Used for training and testing models in standard situations;

[0019] ② A single user performs different movements to verify the performance of the model under different movements. This dataset is named D cm ;

[0020] ③ Recruit a group of users to verify the performance of our model under different users. This dataset is named D cs , the distance and angle of the lateral displacement and annular displacement of the elbow guard are defined as η and β respectively;

[0021] Official implementation:

[0022] Initially, β = 0, η = η min , in textile sensor bending N B After that, if η>η max In the case of , if β>360, the process ends, if η>η max If the situation does not occur, then input η+=Δ η , and repeat the bending N B The action continues until η>η max If β>360 does not occur, input β+=Δ β , and repeat η=η min until the end of the assignment.

[0023] As a preferred solution of the present invention, it also includes the relationship between the sensor resistance and the stretching ratio, wherein there is a nonlinear relationship between the resistance value and the sensor stretching length, and the corresponding relationship also changes with stretching and rebound, and the resistance signal changes with the change of the sensor displacement.

[0024] As a preferred solution of the present invention, the textile sensor is implanted into a common elbow pad, and the sensors are evenly distributed around the elbow in a ring shape.

[0025] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0026] The present invention targets existing methods that mainly classify specific movements rather than accurately reconstruct human movements using flexible sensors, and generally cannot compensate for large displacements. Some methods do not support sensor rotation. Due to the easy deformation characteristics of flexible sensors, it is more difficult to achieve robust motion tracking when the sensor has a position offset. In contrast, the method of the present invention can not only robustly track human joint movements, but also be robust to large annular and lateral device displacements. In addition, since the predictor of the present invention can compensate for displacements through a robust algorithm, the method of the present invention does not require calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0028] Figure 1 A schematic diagram of the structure of the preparation process of the flexibly placed textile sensor for robust joint motion tracking proposed by the present invention;

[0029] Figure 2 A schematic diagram of the experimental process structure of the flexibly placed textile sensor for robust joint motion tracking proposed by the present invention;

[0030] Figure 3 A schematic diagram of a textile sensor structure of a textile sensor for robust joint motion tracking that can be flexibly placed, as proposed by the present invention;

[0031] Figure 4 This is a schematic structural diagram of a situation in which R2 of the textile sensor for robust joint motion tracking that can be flexibly placed is placed on one side of the olecranon bone;

[0032] Figure 5 A schematic diagram of sensor readings of a textile sensor placed at an elbow bent and stretched in a textile sensor for robust joint motion tracking that can be flexibly placed according to the present invention;

[0033] Figure 6 This is a schematic diagram of placing R2 and R3 of the textile sensor for robust joint motion tracking that can be flexibly placed proposed by the present invention on one side of the antecubital fossa of the elbow;

[0034] Figure 7 A schematic diagram of the sensor reading structure of the textile sensor for robust joint motion tracking that can be flexibly placed, wherein R2 and R3 are placed at one side of the antecubital fossa of the elbow;

[0035] Figure 8 This is a schematic diagram of the structure in which R3 and R4 of the textile sensor for robust joint motion tracking that can be flexibly placed according to the present invention are placed on one side of the olecranon bone;

[0036] Fig. 9 A schematic diagram of the sensor reading structure in which R3 and R4 of the textile sensor for robust joint motion tracking that can be flexibly placed according to the present invention are placed under one side of the olecranon bone;

[0037] Fig.10 A schematic diagram of the sensor composition structure of the textile sensor for robust joint motion tracking that can be flexibly placed proposed by the present invention;

[0038] Fig.11 A schematic diagram of the structure of the change of resistance of the sensor with stretching of the textile sensor for robust joint motion tracking that can be flexibly placed proposed by the present invention;

[0039] Fig.12 A schematic diagram of the circuit board design structure of the textile sensor for robust joint motion tracking that can be flexibly placed proposed by the present invention;

[0040] Fig.13 A schematic diagram of the data collection flow structure of the flexibly placed textile sensor for robust joint motion tracking proposed by the present invention;

[0041] Fig.14 A schematic diagram of a single-user average tracking error structure of a textile sensor for robust joint motion tracking that can be flexibly placed and is proposed in the present invention under the condition of a lateral displacement η of the sensor;

[0042] Fig.15 A schematic diagram of a single-user average tracking error structure of a textile sensor for robust joint motion tracking that can be flexibly placed and is proposed in the present invention under the condition of a sensor annular displacement β;

[0043] Fig.16 A schematic diagram of the relationship between the number of data samples of the textile sensor for robust joint motion tracking and the bending angle of the elbow joint proposed by the present invention;

[0044] Fig.17 A schematic diagram of the tracking error structure of the textile sensor for robust joint motion tracking under different elbow bending angles proposed by the present invention that can be flexibly placed;

[0045] Fig.18 A schematic diagram of the tracking error structure of the textile sensor for robust joint motion tracking at different elbow bending speeds that can be flexibly placed according to the present invention;

[0046] Fig.19 A schematic diagram of the average tracking error of different motion type data in the motion data set DCM of the textile sensor that can be flexibly placed for robust joint motion tracking proposed by the present invention and the effect structure before and after incremental learning;

[0047] Fig. 20 Schematic diagram of the average tracking error and arm circumference structure of participants of the flexibly placed textile sensor for robust joint motion tracking proposed by the present invention;

[0048] Fig.21 A schematic diagram of the structure of the average tracking error of the textile sensor for robust joint motion tracking that can be flexibly placed and proposed by the present invention, as a function of the number of sensors;

[0049] Fig. 22 Schematic diagram of the comparative structure of the flexibly placed textile sensor for robust joint motion tracking proposed in the present invention and other learning-based methods.

[0050] Fig.23 Schematic diagram of potential applications of the flexibly placed textile sensor for robust joint motion tracking proposed in the present invention for motor function rehabilitation. DETAILED DESCRIPTION

[0051] In order to more clearly explain and illustrate the technical solution and implementation of the present invention, several preferred specific embodiments for implementing the technical solution of the present invention are introduced below.

[0052] The following description is merely exemplary in nature and is not intended to limit the present disclosure, application, and use. It should be understood that in all of these figures, the same or similar figure numbers indicate the same or similar parts and features. The various figures only schematically represent the concepts and principles of the embodiments of the present disclosure, and do not necessarily show the specific dimensions and proportions of the various embodiments of the present disclosure. Specific parts in specific drawings may use exaggerated methods to illustrate the relevant details or structures of the embodiments of the present disclosure. The various publications, patents, and published patent specifications cited herein are incorporated herein by reference in their entirety. The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention.

[0053] In the description of the present invention, unless otherwise clearly specified and limited, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance; unless otherwise specified or explained, the term "plurality" refers to two or more; the terms "connected" and "fixed" should be understood in a broad sense, for example, "connected" can be a fixed connection or a detachable connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0054] In the description of this specification, it should be understood that the directional words such as "upper", "lower", "left", and "right" described in the embodiments of the present application are described at the angles shown in the accompanying drawings and should not be understood as limitations on the embodiments of the present application. In addition, in the context, it should also be understood that when it is mentioned that an element is connected to another element "upper" or "lower", it can not only be directly connected to another element "upper" or "lower", but also indirectly connected to another element "upper" or "lower" through an intermediate element.

[0055] Test Example 1

[0056] Flexible placement of textile sensors for robust joint motion tracking: Single-user data acquisition experiments:

[0057] ①Preparation:

[0058] A 26-year-old female was invited to prepare a single-user training dataset. (The participant was labeled P1.) After informing the participant of the purpose and procedure of the experiment, the participant agreed to participate in the experiment. The participant's arm circumference was 23 cm to evaluate the wearing experience of the elbow pad and the extension of the sensor (keeping the elbow pad close to the arm so that the sensor can be effectively extended). The expected arm circumference that can comfortably wear the elbow pad is 20-26 cm.

[0059] ②Data collection process:

[0060] 1. Participant P1 entered the motion tracking studio wearing an elbow pad. Reflective markers were added to the elbow pad and tracked by an infrared camera. The Qualisys motion capture system was used to track and calculate the 3D position of the marker at a speed of 300 frames per second. The angle between the centerline vector of the upper arm and the centerline vector of the forearm, that is, the rotation angle of the elbow joint, was calculated from the tracked position. The acquisition time of the signal collected by the sensor and the joint angle collected by the motion capture system was synchronized by triggering two acquisition software at the same time. In view of the different frame rates of the two signal sources, this work will downsample the joint value to ensure the consistency of the sample size. These steps are also implemented in other sampling processes below.

[0061] 2. When collecting D ssWhen the elbow guard is rotated, the range of the elbow guard's annular displacement is β∈[0°, 360°]. Since this range is sampled within 360 degrees, the elbow guard can be highly robust in the case of arbitrary rotational displacement of the elbow guard in subsequent applications. The elbow fossa is defined as the starting position of the elbow guard offset during the acquisition process. This position can also be any point around the joint where the upper arm and lower arm are connected. The range of the elbow guard's lateral displacement is defined as η∈[ηmin, ηmax], where ηmax=-ηmin=4cm. For each trial, the participant needs to repeatedly bend N B = 8 arms. For each bending cycle, participants were required to flex and extend their elbows at different speeds within the range of arm bending (approximately θ∈[40°, 180°]). After each bending trial, the experimenter adjusted the elbow guard to have a small lateral displacement Δη=1 cm relative to its original position until it reached ηmax or ηmin. After sampling each sampling point of the lateral displacement, the experimenter rotated the elbow guard relative to the arm by Δβ=10° and repeated the above procedure until the elbow guard's rotation position covered each rotation sampling point on the arm. The acquisition process took a total of approximately 1.3 hours.

[0062] 3. Collecting multi-motion dataset D cm The participants were required to perform the following four movements: walking, running, clapping, and jumping. Since people often perform these four movements in daily activities, no further pre-training was required for the data collection process. The trained model can be used as a reference for the application of the proposed prototype in daily life. Walking, running, clapping, and jumping are activities that may be performed every day in daily life and require a certain range of motion. Performing these activities will cause the elbow joint to have different rotation angles and movement speeds. During each acquisition process, each type of movement will last for 100 seconds. Participants will repeat the acquisition process under 5 different sensor lateral displacements and circumferential displacements. The offset position is within this range (β∈[0°, 360°], η∈[ηmin, ηmax]), and the entire acquisition process lasts 50 minutes.

[0063] Test Example 2

[0064] Flexible placement of textile sensors for robust joint motion tracking: a multi-user data acquisition experiment:

[0065] ①Preparation:

[0066] Nine users were invited to participate in the collection, including five males: P5, P7, P8, P9, P10, and four females: P2, P3, P4, P6. Their ages ranged from 22 to 25 years old, and their arm circumferences ranged from 20 cm to 26 cm.

[0067] ②Collection process:

[0068] In order to construct the dataset D cs Each participant needs to collect D cm The steps of the dataset were used for data collection. The distribution of reflective markers placed on their bodies was the same. In each trial, each participant performed the above four actions for 40 seconds. Next, the elbow pad was transferred to a random position within the displacement range (β∈[0°, 360°], η∈[ηmin, ηmax]). After that, the collection experiment was repeated three times.

[0069] ③Model training process:

[0070] Training dataset D ss The training set (D TR -80%) and the test set (D TR -20%). Sensor readings include 10°-360° spacing, 10° circular displacement, and 1 cm lateral displacement from -4 cm to 4 cm spacing.

[0071] ④LIghtGBM algorithm for motion tracking:

[0072] The problem of estimating the elbow angle is a regression problem. To solve this problem, this work uses a LightGBM model, which is essentially a gradient boosting decision tree (GBDT) algorithm. By using the gradient-based single-edge gradient sampling (GOSS) algorithm and the exclusive feature bundling (EFB) algorithm, LightGBM increases the training speed of the traditional GBDT by more than 20 times while maintaining almost the same accuracy. Compared with other methods, GBDT uses decision trees to learn the input space X s To the gradient space 's mapping.

[0073] The decision tree splits each node with the largest information gain. Let {x1, ..., xn} represent the S-dimensional input vector, and let {g1, ..., gn} represent the negative gradient of each iteration, then the information gain is defined as follows:

[0074]

[0075] Where n o =∑I[x i ∈o], In order to retain all instances with larger gradients and randomly sample instances with smaller gradients, the GOSS algorithm sorts the training data according to the gradient. Then, it makes the gradient maximum-a×100% training data with larger gradients and randomly selects (1-a)×100% training data with smaller gradients to calculate the estimated variance gain:

[0076]

[0077] whereA l ={x i ∈A:x ij ≤d},A r ={x i ∈A:x ij >d},B l ={x i ∈B:x ij ≤d}, B r ={x i ∈B:x ij >d}

[0078] LightGBM also implements the exclusive feature bundling (EFB) algorithm to reduce the number of features. For high-dimensional, sparse features, the EFB algorithm reduces the calculation of feature values ​​​​to be zero by bundling mutually exclusive features together to form a single feature package. The specific parameter settings are obtained by Grid Search. If the tracking error does not improve within ten rounds, the model will stop training. In order to retain all instances with larger gradients and randomly sample instances with smaller gradients, the GOSS algorithm sorts the training data according to the gradient.

[0079] ⑤Incremental learning:

[0080] In order to ensure accurate tracking of different sensor offset positions and motion types on the same user and new users, incremental learning is performed. Incremental learning enables the LightGBM algorithm to accommodate new data, including clapping, running, walking, jumping, and new user data, and give the corresponding arm joint angles. In addition, the algorithm does not need to access previously used data in the subsequent incremental learning process. It can learn additional information from new data without catastrophic forgetting. Incremental learning only requires input of new data once, without repeated calibration like IMU. The work expects that when encountering new actions or users that are not in the training dataset, the knowledge base can be quickly updated and accurately predict their elbow joint angles in real time. When a user wants to perform an action that is not included in the DSS dataset, or the system needs to predict the elbow joint angle of a new user, he / she needs to wear a smart elbow guard and collect resistance data and corresponding optical angles for 100 seconds. In order to test the effect of incremental learning, the experimenters conducted experiments on the DCM and DCS datasets. The experiment only uses the first 2000 data (which only takes 100 seconds to collect) for incremental learning, and tests the remaining data.

[0081] Experimental Example 3

[0082] Experimental results:

[0083] ①Dataset and training results:

[0084] Last D ss D cm , and D cs The datasets contain 9529232308 and 13841 samples respectively. From the data distribution of the original training dataset, it can be observed that a large number of samples are distributed in the interval of the circular displacement angle β = [160, 180]. This largely corresponds to the initial static posture, that is, the moment when the subject fully stretches his / her arms. In order to balance the data, this work reduces these data samples to 1000 before training. This work also tests the case of training LightGBM with the full dataset without using the balancing strategy. However, the model overfits the samples in this range. During the training phase, this work overfits the samples in the training set D TR The model was optimized on the test set D TR The average tracking error is calculated on the test set D after 3.46 seconds of training. TR The average tracking error of the human elbow joint is 2.54 degrees, and its lateral displacement and circular displacement have no obvious effect on the tracking error. This verifies that the method of the present invention can be applied to various lateral and circular displacements, and illustrates the robustness of the algorithm of this work.

[0085] In exploring D TE The relationship between the joint motion angle and motion speed and the prediction error in the data set, and the influence of the joint motion angle and motion speed on the prediction error are plotted respectively. Fig.15 The number of samples at different joint bending angles is shown. The data samples are mainly concentrated in the elbow bending angle range of [70, 90] and [150, 170]. In contrast, there are fewer data samples around 130 degrees. Fig.15 It can be seen that the maximum and minimum tracking errors are located near the joint bending angles of 30 degrees and 160 degrees, respectively. The smallest tracking error may be due to the large number of data samples distributed near the joint bending angle of 160 degrees. The largest tracking error may be caused by the small number of data samples distributed in the joint bending angle interval [20, 30]. According to statistics, the correlation value between the number of samples and the average error is -0.46. This means that they are weakly correlated, that is, increasing the number of samples may lead to a slight trend of decreasing tracking error. Therefore, the number of samples at a specific joint angle in the dataset may have an impact on the error. The speed of elbow bending movement is calculated by dividing the change in bending angle by the time taken. In addition, this work also conducted a correlation analysis between movement speed and average tracking error, and the result was 0.02, indicating that there is no correlation between movement speed and average tracking error, that is, movement speed has no obvious effect on tracking error.

[0086] ②Tracking results for different types of motion:

[0087] The data collection process for different types of motion is as follows: Fig. 20 As shown. Among them, the experiment also compared the tracking errors with and without incremental learning. It can be seen that without incremental learning, the average tracking error of clapping is the highest, followed by running, and then the average tracking error of walking and jumping exceeds 8 degrees, and the interpolation of the upper and lower limits of the box plot exceeds 20 degrees. In the case of incremental learning, the difference between the four types of movement is smaller. In addition, the difference between the average tracking error and the upper and lower limits of the box plot has also decreased. The upper limit is below 16 degrees, and the average tracking error is below 8 degrees. The tracking errors of the four types of movement have dropped significantly before and after incremental learning. Since the data do not follow a normal distribution, this work uses the Mann-Whitney U test to test whether the tracking errors before and after incremental learning are statistically significant. The results (U=5422459.50, P<0.001) show that there is a significant difference in the average tracking error before and after incremental learning. This proves that incremental learning is suitable for predicting different types of movement and reducing the difference in tracking errors of different types of movement. This experiment also verifies that for the dataset D cm , the average tracking error of our method is 2.93 degrees.

[0088] ③Tracking results for different users:

[0089] Training dataset D created by P1 ss In order to test the cross-user prediction error, we use the cross-user dataset D cs The tracking performance of the smart elbow brace for other users was measured above. Without incremental learning, the tracking errors between the 10 users were significantly different. However, the tracking error and error range were significantly reduced after incremental learning. The Mann-Whitney U test also confirmed that there was a significant difference in the error before and after incremental learning (U=213823.00, P<0.001). A conclusion was drawn: the difference in tracking error between users was very small. The tracking error for different users was less than 16 degrees, and in dataset D cs The average tracking error on the upper arm is 4.21 degrees. This shows the robustness of the method for tracking different users. In order to explore the possible relationship between user arm circumference and tracking error, this work measured the average arm circumference of the corresponding users (the average of the upper and lower arm circumferences). Since the differences between different users are small, the user's arm circumference has no significant effect on the tracking results. In summary, incremental learning improves the average tracking error and the tracking error range, while arm circumference has no significant effect on the tracking results.

[0090] The sensor readings change as the joint moves and the sensor stretches. Figure 5The sensor readings at different sensor offset positions are shown. It can be seen from the figure that changes in the elbow joint angle will cause significant changes in the readings of 2-3 sensors at the same time. The sensors with significant changes in readings are located in the semicircular area on the side of the olecranon bone of the arm, which is effectively stretched when the joint rotates. In contrast, the sensors on the side of the fossa are not stretched, so their readings do not change significantly. The method of this work can effectively cover the stretched semicircular area and can support the elbow guard to achieve 360 ​​degrees of circular displacement and more than 4 cm of lateral displacement. Therefore, the number of suitable sensors is 4-6; the results of using different numbers of sensors are also compared numerically (as shown in Table 1). From the results in the table, it can be concluded that the accuracy increases with the increase in the number of sensors. However, when the number of sensors is 4 or 5, the growth rate of accuracy is low, and there is no significant error reduction. Therefore, this work chooses to use six sensors to predict the joint angle, which can save costs while maintaining high accuracy.

[0091] ④System delay:

[0092] The total delay of the system is 69ms. Its delay can be divided into three parts: the first part is the sampling interval of the circuit board, which sends data every 50ms; the second part is the process of estimating the joint angle by the lighttgbm model, with a delay of 0.007ms. The third part is the network delay. Since the server and mobile Bluetooth transmitter are connected through a TCP network, the delay depends on the quality of the network signal. In the experiments of this work (conducted using a local wireless area network), the delays of the server-mobile and server-computer stages were 2ms and 17ms, respectively.

[0093] ⑤Comparison with other learning-based methods:

[0094] Without using incremental learning, on dataset D ss The average tracking error of this method is compared with other nonlinear machine learning methods and deep neural networks. These methods include XgBoost (eXtreme GradientBoosting), random forest, K-nearest neighbors, and fully-connected neural network.

[0095] XgBoost is a scalable end-to-end boosting tree system. Random Forest is a common and effective algorithm in the field of machine learning. The KNN algorithm is a non-parametric classification and regression method. FCN contains 8 layers (512 units per layer). They are all popular methods in the field of machine learning. It can be observed that the average tracking error and variance of XgBoost, Random Forest FCN and the method of this work are relatively close, while the method of this work requires less training time and running time. KNN takes less time. But its average error and variance are greater than other methods. In general, if this work wants to make a trade-off between time consumption and tracking accuracy, LightGBM is the best machine learning method.

[0096] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for tracking elbow joint motion using a flexibly placed textile sensor, wherein the textile sensor is composed of a standard textile elbow guard and a sparse network of six flexible stretch sensors, wherein the preparation process of the textile sensor is as follows: S1: winding silk fibers into yarn; S2: Based on S1, the polyurethane solution is diluted to 6.5 mg / ml, mixed with 5 mg / ml silver nanowire solution in equal volume ratio, and the nanosilver wire solution with a concentration of 5.2 mg / ml is poured on the surface of the woven yarn structure, immersed and dried to obtain conductive silver wire; S3: Based on S2, the head of the conductive silver wire is fixed to the end of the textile sensor by sewing; S4: Based on S3, wrap the end of the conductive silver wire around the head of the conductor and wrap it with insulating glue; S5: Based on S4, the end of the conductor is directly soldered to the six-channel PCB board and sealed with hot melt adhesive. When the flexible stretch sensor is stretched, the conductive fibers separate and the contact area decreases, resulting in an increase in resistance value. The flexible stretch sensor is manually sewn into the elbow pad using a flat needle; including the process of collecting data from the textile sensor with a six-channel PCB board: The Wheatstone bridge structure is used to calculate the difference between the voltage of each flexible sensor and the reference voltage Vref = Vcc / 2, where Vcc represents the voltage connected to the circuit. A pair of identical resistor units is also used to obtain the reference voltage. The input voltage V adc c cin is defined as: in, R i Represents the voltage divider resistor, Rsensor i represents the i-th flexible sensor, Gain represents the amplification factor of the amplifier; The voltage measurement result is processed by a low-pass filter with a bandwidth of 300 Hz and converted into digital form within [0, 1023]. The six-channel PCB board sends the digitized sensor signal to the mobile terminal via Bluetooth, and then it is further transmitted to the remote server. The server executes the calculation process of the LightGBM model to obtain the joint angle of the elbow joint. The data is sent at a speed of 20 frames per second through the Bluetooth chip. The channel is selected in turn by the multiplexing divider. The server is configured with a single-core CPU, 2GB of memory, and the operating system is Ubuntu16.0464 bit; it also includes the data set acquisition process: The data set collection process is divided into three stages: ① A single user performs activities at different device offset positions, and the collected data set D ss Used for training and testing models in standard situations; ② A single user performs different movements, and the collected data set is named D cm , to verify the performance of the model under different motion conditions; ③ Data collection of a group of users, each user collects data set D cm The data set is named D cs , to verify the performance of the model under different users; The three phases include the following steps: Initially, β = 0, η = η min , in textile sensor bending N B After that, if η>η max In the case of , if β>360, the process ends, if η>η max The situation does not occur, then input η+=Δ η , and repeat the bending N B The action continues until η>η max If β>360 does not occur, input β+=Δ β , and repeat η=η min The operation is continued until the end, wherein the distance and angle of the lateral displacement and annular displacement of the elbow guard are defined as η and β, respectively. max is the maximum distance of lateral displacement, η min is the minimum value of the lateral displacement distance.

2. The method for tracking elbow joint motion using a flexibly placed textile sensor according to claim 1, characterized in that: It also includes the relationship between the resistance of the flexible stretch sensor and the stretch ratio, where there is a nonlinear relationship between the resistance value and the stretch length of the flexible stretch sensor, and the corresponding relationship also changes with stretching and rebound, and the resistance signal changes with the displacement of the flexible stretch sensor.

3. The method for tracking elbow joint motion using a flexibly placed textile sensor according to claim 1, characterized in that: The textile sensor is implanted on a common elbow brace, and the flexible stretch sensors are evenly distributed around the elbow in a ring shape.

Images