Neural Comput & Applic DOI 10.

1007/s00521-012-1045-8

ORIGINAL ARTICLE

Hierarchical projection regression for online estimation of elbow joint angle using EMG signals
Yang Chen Xingang Zhao Jianda Han

Received: 16 May 2011 / Accepted: 22 June 2012 Springer-Verlag London Limited 2012

Abstract Exoskeletons offer great facilities to the elderly and disabled people with respect to extending their moving ranges and reacting to certain physical activities. Electromyogram (EMG) signals, which are derived from the neuromuscular system, provide an important access to the human-robot interface. On one hand, EMG signals can be used for real-time estimation of the motion intention of human body, e.g., the current joint angle status. On the other hand, however, the process of the mass EMG data, which are captured instantaneously from skin surface, challenges the state-of-the-art technology. Because its nonstationary and randomness, it is difcult to extract the valuable and stable features from the raw EMG signals. This paper investigates into the learning process of high dimensional EMG signals with a hierarchical mechanism that projects the original data into a lower feature space to achieve a local rened mapping from the EMG signals to the motion states of the human body. This hierarchically projected regression algorithm constructs incrementally a tree-based knowledge library, whose components represent local regression models. The components will be retrieved online efciently and contribute to the estimation of the

motion states. A great number of experiments are carried out to evaluate the accuracy of this novel algorithm. Keywords Electromyogram (EMG) signals Exoskeletons Hierarchical projection regression k-Means clustering Linear discriminant analysis

1 Introduction Exoskeletons, as a kind of wearable robot, recently get extensively researched and applied in the domain of physical assistive support in home, rehabilitation in medical eld and stress powering in military. Whatever the role of wearable robot is, it is necessary for exoskeleton robots to interpret the intention of human motion in advance of the actual movement [1]. The electromyogram (EMG) signal, generated from the surface muscles contraction, can be directly detected through sensors attached to the surface skin of the users. In this scenario, EMG signals provide rich information in real time that can serve as the interface of the humanmachine system. By using EMG signals, wearable robots achieve desired exibility because the inherent consciousness of human can be maintained to the utmost extent. However, EMG signal is a kind of bio-signal which is very weak and can probably be affected by various noises. One of the greatest challenges in understanding EMG signals is their non-stationary and nonlinear characteristics. The EMG features, extracted from the raw signals, should be steady and valuable when they serve as the commands of exoskeletons, although the extracting process is intractable. Raw signals are usually sampled at a constant frequency. Therefore, the human motion estimation becomes a difcult issue due to its high dimensional inputs, i.e., a

Y. Chen (&) School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, China e-mail: chenyag@gmail.com X. Zhao J. Han State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang 110016, China e-mail: zhaoxingang@sia.cn J. Han e-mail: jdhan@sia.cn

123

Neural Comput & Applic

segment of time series. For example, the muscles are often modeled rstly by identifying several parameters before the actual modeling proceeds, such as Huxley model and Hill model for muscle activation and joint movements [26]. Many researchers analyzed the relations between the joint movement and the EMG statistics, rst or second order [7], and other features. Their studies show the strong nonlinearity of the relations. Based on the obtained features, the manipulators can be controlled to move according to the muscle movement of the users and enable to support the daily activities. For example, Kiguchi et al. [8] and Sato and Yagi [9] developed an exoskeleton for human upper-limb motion assistance based on EMG signal. In their work, root mean square is extracted to serve as the control command straightaway. However, the underlying ve feature knowledge is poorly explained by this na because a large amount of information is ignored. To remedy this, several nonlinear regression formulations have been proposed and tested [5, 10, 11]. Some latest methods are Gaussian process regression [12], which can learn the inverse dynamics of a seven degrees-of-freedom SARCOS anthropomorphic robot arm, and Bayesian regression [13], which can be used to predict EMG activities from neural ring. Unfortunately, the above traditional methods did not discuss the variety of the subjects, and their accuracy is unsatisfactory when it comes to estimate the continuous motion. There are some researchers who have tried to map the EMG signals directly to the output, i.e., the joint angle states, without the steps of modeling and parameter identication. These are data-driven approaches, such as Articial Neural Network (ANN). ANN is one of the black-box models dealing with mapping from multi-inputs to multioutputs, which is a combination of many nonlinear functions. Wang and Buchanan [5] designed a four-layer neural network that includes two hidden layers, to estimate the activation of ten muscles and proposed an algorithm named adjusted back-propagation to conquer the immeasurability of supervised information. Gopura and Kiguchi [11] proposed a fuzzy-neuron mapping from EMG signal to joint torque. Bu et al. [14] used a neural network to predict the movement and its occurrence probability. However, the prediction accuracy depends on the context information of a task. In general, ANN is not satisfactory in terms of realtime processing, especially for high dimensional training dataset. It cannot be applied in conditions where online training and updating are required. Another black-box state-space model was proposed in [15] for joint angle estimation based on EMG signal. This method needs position tracking system for tracking the joint movement to avoid estimation drifting. The method proposed in this paper, named hierarchical projected regression (HPR), focuses on the high dimensional

EMG data learning. By using HPR, the joint angle of human elbow can be estimated online. During the training stage, the HPR algorithm constructs a tree-based knowledge library, whose components represent local regression models. When it comes to the real-time operation stage, the joint angle can be retrieved online from the HPR tree using the EMG signals. At both the training stage and the real-time operation stage, the EMG signal is recorded and a sliding time-window is used to infer the stable features. In addition, in the training phase, the real angle of the human elbow is measured by an inertial measurement unit (IMU). The remainder of the paper is arranged as follows. Experimental conguration and data acquisition are described in Sect. 2. Section 3 presents hierarchical projected regression algorithm in details. Section 4 describes the experimental results and the corresponding analysis. Section 5 discusses some extensions. A concise conclusion is drawn in Sect. 6.

2 Experimental conguration and data acquisition Experimental conguration for joint angle estimation online is described in Fig. 1. We obtain raw EMG signal from the A/D card rst. After the data are preprocessed, the signals will be fed into the HPR, and then, we obtain the joint angle online. The estimated angle can be sent to the exoskeletons robot through the UDP protocol. Finally, robot will move following the human elbow motion. 2.1 Samples for training In this paper, we try to recognize the elbow joint motion. Because motions result from contraction of the muscles in the forearm, we use only one channel of surface electrodes to measure EMG signals from the bicep which is one of the muscles mostly responsible for elbow joint motions. The EMG signals are sampled with the A/D card (ADVANTECH PCI-1716) and tri-electrode amplier device (MyoScan) with the frequency of 1 kHz, as shown in Fig. 2a. An angle sensor (VECTORNAV VN-100), which is mounted on the subjects elbow, supplies the joint angles as the supervising signals. Raw EMG signals and the corresponding elbow joint angle are seen in Fig. 2b. First of

EMG

Preprocess HPR

Robot

UDP

Fig. 1 Experimental conguration for joint angle estimation online

123

Neural Comput & Applic

(a)
EMG signal

Raw signals

High-pass filter (20Hz)

Notch filter (50Hz)

IMU

Low-pass filter (3Hz)

Rectification

Amplifier

Normalization (loads)

Sliding window

EMG samples

(b)
Joint Angle (deg.) raw EMG signals
0.5 0 -0.5 0 2 4 6 8 10 12 14 16

Fig. 3 Preprocess of raw signals. This process ends up with the envelope of the rectied and normalized EMG signal when it ows into the sliding window for the nal samples

100 50

branch tree. Assume that the training dataset is S = {(xi, yi)|i = 1,2,,N, where xi [ Rd91 represents EMG signals corresponding to the known joint angle yi. Here, both xi and yi are preprocessed. The following subsections will discuss how to construct the HPR tree. 3.1 Hierarchical issues

0 0 2 4 6 8 10 12 14 16

Time (s)
Fig. 2 Sampling the EMG signals. a Experimental setup. b Raw EMG signals and the corresponding elbow joint angle

all, the subject initially grasps the loads and keeps the arm at perpendicular state. Further, the subject exes the arm to the high position and then ex-exes to the original position. This testing process repeats several times with different loads, while EMG signals and joint angles are recorded. 2.2 Data ltering and sliding window Raw signals should be fed to the digital lters before the HPR tree is trained. A set of lters are used, including a high-pass lter (cutoff frequency 20 Hz) to remove any DC offsets or low frequency noise, a notch lter (50 Hz), a full wave rectication and a low-pass lter (cutoff frequency 3 Hz) [13], as shown in Fig. 3. The ltered data are then normalized with variance of one and means of zero. All the lters used are a 2-order Butterworth lter. The width of the data window is 128 ms, and the sliding window rolls forward in a step of 32 ms.

3 Algorithm descriptions Hierarchical projected regression is a coarse-to-ne method with a hierarchical formulation, known as multi-

Hierarchical projected regression tries to t the training data with micro regression models. The process of selecting a specic micro model, given the real-time EMG signal as input, can be described by a sequential decision-making process corresponding to the traversal of the HPR tree. The statistics of those micro models are tested as the traversal criterion iteratively. Particularly, the variance and the sample quantity in each model are used to evaluate the tting performance. When the variance and the quantity are less than certain thresholds, the micro model will be considered as a leaf node of the HPR tree. Otherwise, the node is an inner one and should be split by calling the clustering algorithm which will be discussed later. Motivated by the nite mixture models [16], the hierarchical issue we proposed is able to approximate the distribution of the EMG signals. Some existing methods, such as Gaussian Mixed Models (GMMs), are also the nite mixture models. But it is assumed that all the samples in GMMs can be approximated by the Gaussian distribution. There is no effective theory supporting how many models should be selected for GMMs, and it often depends on the context and the users experience. To overcome the defect, HPR relaxes this constraint and only sets the upper bound of the sub-node quantity in each level of the tree considering the efciency while allowing the micro models to be generated automatically. After training, the nal quantity can be adjusted according to the size of the training set. Therefore, this adjustment leads to a much more elaborate model.

123

Neural Comput & Applic Fig. 4 The prole of the hierarchical projected regression tree. The number represents the nodes ID of the HPR tree. NO. 0 is the root of the tree. There are ve inner nodes which include 1, 5, 7, 12 and 14
0

12

7 2 3 4 6 11 13

14 18

8 9 10

15 16 17

Besides, the hierarchical framework brings in an advantage of parallel retrieval strategy. Hierarchy supplies an access to retrieve any individual observation sequentially. The terminals of the tree, i.e., the leaves of the tree, represent the local models which usually present excellent aggregation. When retrieval procedure starts, the branches of HPR are traversed until the leaf node is accessed. Then, a traditional regression method, k-nearest neighbor regression, is called to nd the nearest point as the outcome. A nal HPR tree that shown in Fig. 4 is constructed on 1,579 samples. 3.2 k-Means clustering In the rst half stage, the joint angle regression of HPR is treated as one of the classication problems, although the sample pairs of EMG signals have no class labels. Practically, some clustering algorithms can be used to generate the pseudo labels for the continuous EMG signals. We particularly cluster the real angle, i.e., yi, into K classes. Then, the ID of each class is named as the labels of the corresponding EMG signals, xi. There has been a great number of clustering algorithms. k-Means algorithm is one of the fastest algorithms and can run dynamically in the manner of iterative renement. This approach partitions all observations into K clusters in which each observation belongs to the cluster with the nearest mean. k-Means formulation is arg min
S K X X yj vi 2 i 1 yj 2 S i

yi, share the same class label. k-Means algorithm is ready to converge to a local minimum. Initial centers decide whether the algorithm can get to the global optimal. Consequently, the classication can be improved by selecting the appropriate cluster centers. 3.3 Feature extraction with linear discriminant analysis After the labels are obtained, EMG signals can be projected into lower subspace. The regression performance can be improved because the uncertainty brought by the raw signals will be reduced in lower space. Assume that the projecting matrix is A [ Rd9q. The projected samples are represented as zi AT xi 2

1 X s:t: vi yj N i y 2S
j i

where Si denotes the subsets and Ni represents the sample quantity of the ith subset. Both parts of each subset, xi and

where zi represents the projected sample. There are many methods to solve the projecting matrix for the dimension reduction purpose because of q \ d. Principle component analysis (PCA) is one of the mostly used methods for dimension reduction. This approach chooses a projecting matrix in such a way that it minimizes the average projection cost, dened as the mean squared distance between the samples and their projections [17, 18]. Unfortunately, it could not retain the class information. Linear discriminant analysis (LDA) is another method for dimension reduction. It is generalized from the Fisher discriminant criterion for two-cluster classication problem, which seeks for an embedding transformation so that the between-class scatter is maximized and the within-class scatter is minimized. LDA is a traditional but useful method for dimensionality reduction. However, it is also a linear transformation. For those nonlinear circumstances, the classication accuracy of LDA is usually much lower. This paper proposes an iterative process for the optimization of the projected subspace based on the basic LDA algorithm.

123

Neural Comput & Applic Fig. 5 Dataset with 1,579 samples is divided into three subsets in a lower space by the dimension reduction in LDA criterion. The clusters are marked as c0, c1, c2. ID = 0 represents that the node number is 0, i.e., the root. a The beginning of the Iterative LDA produces the accuracy of 0.57921, which stands the accuracy of the standard LDA formulation. bd After iterations in our conguration, the classication accuracy improves to 0.87852

(a) x 10 -7 ID=0,iter=1,accu=0.57921
2
c0 c1 c2

(b)
2

x 10

-7

ID=0,iter=2,accu=0.8685
c0

1 0

c1 c2

y
-1 -2 -3 -2

-1

-2 -2

-1

-1

x 10

-7

x 10

-7

(c) x 10-7 ID=0,iter=6,accu=0.87602

2
c0

(d) x 10
5

-7

ID=0,iter=9,accu=0.87852
c0 c1

1 0

c1 c2

c2

-1 -5 -2 -3 -1 -10 -2

-0.5

0.5

1.5

-1

x 10

-6

x 10

-7

The standard LDA algorithm is described in the following. The formulation of the standard LDA, described as (3), solves the general eigenvalue problem, Sba = kSwa [17]. Sw and Sb are dened as within-cluster scatter matrix and between-cluster scatter matrix, respectively. Here, a is the eigenvector which represents a column of the matrix A and k is the eigenvalue corresponding to a. det AT Sb A T max : A det A Sw A 8 K X X > > > x ui x ui T S > < w
i1 x2Si K > X > > > Ni ui uui uT : Sb i1

lower space. In Fig. 5a, let K = 3 and we projected the EMG signals into two dimensions. It is obvious that the classication bound of standard LDA is ambiguous, as shown in Fig. 5a. 3.4 Optimization of the projection An iterative strategy is proposed in HPR to improve the classication accuracy. The features in lower dimensional space are used to update the labels and the centers of former space. k-Means algorithm is recalled by HPR which is initialized repeatedly by the updated centers. Once HPR runs for projected optimization, the rened classication that obtains the best projecting matrix will be recorded. For example, the node No. 0 improves the accuracy from 0.57921 of iter = 1 to 0.8685 of iter = 2, as shown in Fig. 5. The projected space and its contour update automatically and change with the iterations. This iteration optimization constructs on general distance between the EMG features and the centers in lower dimensional space. For instance, if the lower dimensional features of xi are represented as zi, the new label of xi is calculated based on zi. Then, the class which has the least distance from its center to zi is chosen as the one that belongs to. The general distance has alternative denitions depending on

1X x N x 2S

If the dataset is partitioned into K clusters, LDA algorithm will obtain most K - 1 nonzero eigenvalue. Standard LDA algorithm obtains a solution of (3) for the dataset which has a unique modal. However, the real EMG signals exhibit multi-modality. The Gaussian assumption in the dataset often leads to the poor classication even in the projected

123

Neural Comput & Applic

Step 1: Initialize the labels randomly for EMG samples. Step 2: Joint angle data are clustered by k-means algorithm. The EMG data are allocated with the corresponding k labels. Step 3: According to the labels, EMG data are divided into k partitions. Using the LDA algorithm, a projecting matrix, A, can be derived by solving the generalized eigenvalue problem. Step 4: Calculate the lower dimensional features, z=ATx, as well as the lower dimensional centers. Step 5: Update the labels of EMG samples employing the lower dimensional features. If the labels maintain unchanged or the iteration gets maximum, jump out of the iteration. Otherwise, go to Step 6. Step 6: Update the centers for the initialization of k-means algorithm, and then return to Step 2.

The manipulator is driven in the velocity mode which has the maximum velocity of 35 per second. 4.1 Experimental results of two kinds of loads Two groups of experiments were carried out in this study. As shown in Table 1, the estimated error for subject A is about 8.39 and 11.21. After many experiments on subject A, a standard deviation of the error, root mean squared error (RMSE), is found to be very small as shown in Table 1. It is obvious that the error of subject B is a bit greater than that of subject A. It is interpretable because the HPR tree is built on the dataset of subject A while relative information about subject B is unknown. EMG signals of different subjects generally resulted in different features. However, the experiments have demonstrated the fact that HPR reduces these diversities to an acceptable range. Figure 8 shows the experimental scenario with subject A. It is found in Fig. 9 that estimated angles by HPR are much closer to the angles that measured by IMU. Particularly, the accuracy of 1.54 kg loads is higher than that of 1.92 kg loads. This is reasonable because the library employed to estimate the joint angle online is created on the 1.54 kg loads. Figure 10 shows the scenario with subject B. Compared to the results of subject B, the results of subject A have higher accuracy. 4.2 Experimental results of drinking task A classical task of drinking tea is also carried out for evaluating the efciency and the robustness of HPR algorithm. The major function of drinking task is exing and ex-exing the elbow, as illustrated in Fig. 11. Figure 11af demonstrate the process of exing the elbow and the cup moving to the mouth. Figure 11gl demonstrate the course of ex-exing the elbow and the cup moving to the table. Meanwhile, the robotic arm moved with the users elbow, as shown in each picture. From these pictures, the arm could track users movement steadily. Figure 12 shows the corresponding curve when the drinking task was executed. There exists a little time delay between the users movement and the estimated result, which is probably caused by the teleoperation of the manipulators and the process of EMG signals.

Fig. 6 Pseudo codes for generating a node

the sample quantity in the node, as shown in (6). If there are enough samples for distance estimation, i.e., Nj [ d for all j, the distance is ready to take the Gaussian probability density as the distance. Otherwise, the Euclidean distance denition is used. Dzi ; cj expfzi cj T Mzi cj g ( 1 R j ; 8Nj [ d ; j 1; 2; . . .; K Mj I; otherwise ci A T u i This approach is different from Incremental Hierarchical Discriminant Regression (IHDR) [19] which only combines the centers of each cluster as a substitution of A and is never optimal. Figure 6 shows the pseudo codes about how to get an individual node for HPR. Figure 7 illustrates the leaves space of HPR. It is obvious that the leaves have been clustered signicantly as well as separated from each other. 6

4 Experiments In this section, all the reported experiments of motion estimation, according to the ethical standards, were carried out in order to evaluate the regression performance of the proposed method. Two subjects, A and B, participated in the experiments. First, subject A performs four trials with the loads of 1.54 kg. Data of these trials were used to train HPR. Then, both A and B were assigned to test our algorithm online. In the testing stage, they were assigned to move the elbow following the same task as in the training procedures. Five trials with 1.54 and 1.92 kg, respectively, were performed for the evaluating purpose. In our experiments, a manipulator was employed to simulate the human elbow. The estimated angle is transferred to the manipulator as the commands for movement.

5 Discussion of some extensions 5.1 Kernel regression in leaf nodes Some other micro models can be developed in the leaf nodes of HPR. For example, the kernel regression on

123

Neural Comput & Applic Fig. 7 The distribution in projection space of inner nodes of HPR tree. The HPR tree refers to Fig. 4
-7

4 2

x 10

ID=0,accu=0.87477
1

x 10

-7

ID=1,accu=0.93121

0.5 0 -2 -4 -0.5 -6 -8 -3 -1 -2

-2

-1

x
x 10
-7

x 10

-7

x
x 10
-7

x 10

-7

ID=5,accu=0.83604
0 -1

ID=7,accu=0.77288

15

10 -2 5 -3 -4 0 -5 -5 -2 -6 -4

-1

-2

x
x 10
-7

x 10

-7

x
x 10
-8

x 10

-8

ID=12,accu=0.78251
4 2 0

ID=14,accu=0.68605

y
-2 -4 -6 -4 -1 0 1 2

-1

-2 -2

-2

x 10

-7

x 10

-8

Table 1 Estimation error of different loads Loads Error () 1.54 kg Subject A Subject B 8.39 0.26 11.21 1.92 kg 11.21 0.38 13.28

5.2 Incremental learning issue With respect to an incremental learning system, the samples come in sequentially in time. The function of updating with the changing environment is critical. As introduced in this paper, the observation of the changing environment is often segmented as a snapshot of world. As a tree-based algorithm, hierarchical projected regression has the advantage of learning incrementally. Those micro regression models are rather well suited for incremental and real-time learning. From the view of HPR, the EMG signals can feed incrementally, and

the local data of each micro model is expected to improve the estimation accuracy, while the complexity increases slightly.

123

Neural Comput & Applic

Fig. 8 The joint angle estimation online for subject A who participated in both the training and testing experiments. The motion of the elbow was estimated online. Meanwhile, the estimation was implemented by the manipulator that is mounted on a platform

Fig. 9 The joint angle estimation online for subject A who participated in both the training and testing experiments. Thin blue curve represents the real joint angle which was captured by the IMU. The thick red curve represents the estimation by the HPR. Two different loads, 1.54 and 1.92 kg, were tested with the same speeds (color gure online)

100

100

Joint angle (deg)

Joint angle (deg)

80

80

60

60

40

40

20

20

10

20

30

10

20

30

Time (s). loads=1.54kg

Time (s). loads=1.92kg

Fig. 10 The joint angle estimation online for subject B who only participated in the testing experiments. Thin blue curve represents the real joint angle which was captured by the IMU. The thick red curve represents the estimation by the HPR. Two different loads, 1.54 and 1.92 kg, were tested at the same speed (color gure online)

100

100

Joint angle (deg)

60 40 20 0

Joint angle (deg)

0 10 20 30

80

80 60 40 20 0

10

20

30

Time (s). loads=1.54kg

Time (s). loads=1.92kg

123

Neural Comput & Applic

Fig. 11 A drinking task was carried out to evaluate the HPR. The user took the cup to his chest and then returned it back to the table. The motion of the elbow was estimated online. Meanwhile, the estimation was implemented by the manipulator that was mounted on a platform

then, the corresponding branches are traversed and updated according to the training data. As a matter of fact, HPR updates much faster than constructing the entire tree. When the new data comes, HPR only needs to search the end model and add a sample in the leaf node. Of course, the leaf node is possible to call spawning procedure as long as the corresponding condition is met. This brings benet to incremental learning for those massive dataset training. 5.3 High dimensional output for multi-joints control Although the experiments of estimating the elbow motion focus on the single output, HPR can be applied in the case of multiple joint motion estimation easily. As far as the

multiple outputs are concerned, the estimated space would be multiple dimensional. When HPR is established, multidimensional dataset is to be clustered with the k-Means algorithm. During the estimation phase online, EMG signals are fed to HPR, and the multi-dimensional motions can be retrieved from the corresponding leaf node as the outcome.

6 Conclusions Electromyogram signals are extensively used in the eld of wearable robots. Because of the non-stationary and uncertain natures of EMG signals, it is a challenging task to extract the valid and steady features from the raw signals.

123

Neural Comput & Applic

80 Real HPR 60

References
1. Bueno L, Brunetti F, Frizera A, Pons JL (2008) Chapter 4: Humanrobot cognitive interaction. In: Pons JL (ed) Wearable robots: biomechatronic exoskeletons. John Wiley & Sons, England 2. Cavallaro EE, Rosen J, Perry JC, Burns S (2006) Real-time myoprocessors for a neural controlled powered exoskeleton arm. IEEE Trans Biomed Eng 53(11):23872396 3. Buchanan TS, Lloyd DG, Manal K, Besier TF (2004) Neuromusculoskeletal modeling: estimation of muscle forces and joint moments and movements from measurements of neural command. J Appl Biomech 20(4):367395 4. Buchanan TS, Lloyd DG, Manal K, Besier TF (2005) Estimation of muscle forces and joint moments using a forward-inverse dynamics model. Med Sci Sports Exerc 37(11):19111916 5. Wang L, Buchanan TS (2002) Prediction of joint moments using a neural network model of muscle activations from EMG signals. IEEE Trans Neural Syst Rehabil Eng 10(1):3037 6. Xu JX, Zhang Y, Pang YJ (2010) A nonlinear parametric identication method for biceps muscle model by using iterative learning approach. 2010 8th IEEE international conference on control and automation, pp 252257 7. Bida O, Rancourt D, Clancy EA (2005) Electromyogram amplitude estimation and joint torque model performance. In: Proceedings of the IEEE 31st annual northeast bioengineering conference, pp 229230 8. Kiguchi K, Rahman MH, Sasaki M, Teramoto K (2008) Development of a 3DOF mobile exoskeleton robot for human upperlimb motion assist. Rob Auton Syst 56(8):678691 9. Sato M, Yagi E (2011) A study on power assist suit using pneumatic actuators based on calculated retaining torques for lift-up motion. In: Proceedings of the SICE annual conference, pp 628632 10. Song Q, Sun B, Lei J, Gao Z, Yu Y, Liu M, Ge Y (2006) Prediction of human elbow torque from EMG using SVM based on AWR information acquisition platform. In: Proceedings of the 2006 IEEE international conference on information acquisition, pp 12741278 11. Gopura RARC, Kiguchi K (2008) A human forearm and wrist motion assist exoskeleton robot with EMG-based fuzzy-neuro control. In: Proceedings of the 2nd biennial IEEE/RAS-EMBS international conference on biomedical robotics and biomechatronics, pp 550555 12. Rasmussen CE, Williams CKI (2006) Gaussian processes for machine learning, chapter 2. MIT Press, Cambridge, MA 13. Ting J, DSouza A, Vijayakumar S, Schaal S (2010) Efcient learning and feature selection in high-dimensional regression. Neural Comput 22(4):831886 14. Bu N, Okamoto M, Tsuji T (2009) A hybrid motion classication approach for EMG-based human-robot interfaces using Bayesian and neural networks. IEEE Trans Robot 25(3):502511 15. Artemiadis PK, Kyriakopoulos KJ (2007) EMG-based position and force control of a robot arm: application to teleoperation and orthosis. In: Proceeding of IEEE/ASME international conference on advanced intelligent mechatronics, pp 16 16. McLachlan GJ, Peel D (2000) Finite mixture models. Wiley, New York 17. Bishop CM (2006) Pattern recognition and machine learning, chapter 4 and chapter 12. Springer Science ? Business Media, New York 18. Ha KH, Varol HA, Goldfarb M (2011) Volitional control of a prosthetic knee using surface electromyography. IEEE Trans Biomed Eng 58(1):144151 19. Weng J, Hwang W (2007) Incremental hierarchical discriminant regression. IEEE Trans Neural Netw 18(2):397415

Joint angle (deg)

40

20

10

15

20

25

30

Time (s). Drinking task

Fig. 12 The comparison between the real angle and its estimation as the drinking task is carried out. The thin blue curve stands for the users elbow angle that is detected by the IMU. The estimation from HPR is drawn with thick red curves. The average error is about 10.33, which is small as can be see from the live video for drinking task (color gure online)

This paper proposed a hierarchical projected regression method to exploit the human motion intention hidden in the EMG signals. HPR is a tree-based algorithm where each node of the tree represents a cluster that can be separated from its parent node using k-Means approach. The child node will be divided recursively until the statistical criteria of the node are satised. In order to extract optimal lower dimensional features, a projecting matrix is derived in each spawning node. Finally, the leaf nodes of HPR tree are the representation of micro regression models. As the retrieval task takes places, the nearest branches in each level of HPR are traversed consecutively until the leaf node is reached. At the leaf node, the nearest point is found by nearest neighbor searching. Hierarchical projected regression algorithm was validated by several experiments where different subjects with different loads and tasks were taken into consideration, and elbow motion was estimated online. The results have shown that HPR maps the high dimensional EMG signals to the elbow movement with satisfying accuracy. On the other hand, a robotic arm is driven to move following the estimated elbow movement. A great number of experiments have indicated that the algorithm we proposed is effective.
Acknowledgments This work was supported in part by the National Natural Science Foundation of China under grant 61035005 and 61075087, Hubei Provincial Natural Science Foundation of China under grant 2010CDA005, Hubei Provincial Education Department Foundation of China under grant no. Q20111105, Natural Science Foundation of Liaoning Province. The authors owe many thanks to Qichuan Ding, Weiran Cao and Anbin Xiong for valuable discussions. Special thanks go to Cheng Chen who helps much about the experiments platform of this study.

123