EMG Pattern Classication

Project Progress Report Engr 9881 Pattern Analysis and Machine Intelligence B.M.Oscar De Silva 201070588 Faculty of Engineering and Applied Science Memorial University of Newfoundland Winter 2011 submitted to: George Mann & Jeannette Byrne
Abstract This report documents the progress of the course project for Engr 9881. The project attempts to implement a machine learning algorithm for identifying user manipulation tasks using only Electro myography signals as the input source. Results for a Linear Discriminant Analysis based classier is reported. This will further be compared with support vector machine and neural network based classiers in future studies. KEY INDEXING TERMS: Upper limb prosthetics, Myoelectric control. Pattern Recognition, Linear Discriminant Analysis

Introduction

EMG based control strategies are extensively used for powered upper limb prosthetics. It has improved from simple threshold control strategies, to more complex pattern recognition based strategies. Pattern recognition techniques was rst proposed by Hudgins et al.[1] using Multilayer Perceptron learning algorithm, utilizing the transient EMG signal. This was later developed in to continuous classication using steady state myoelectric signals [2]. Together with novel surgical procedures [3], this has enabled multifunctional control of upper limb prosthetics (i.e. controlling higher number of degrees of freedom using a limited set of input signals). Dierent Pattern recognition algorithms are appreciated in literature for the purpose. Table 1 summarizes some of the methods and their performances.

Reference Hudgins et al.[1]

Year 1993

Chan et al.[4]

2000

Englehart et al.[5]

2002

Huang et al.[6]

2005

Kuiken et al.[3]

2009

Feature set/Segmentation TD 5 feature set,40ms 5 segments of onset signal TD 4 feature set,40ms 6 segments of onset signal WPT feature set with PCA,256ms 50% overlap continuous on steady state signal TD+AR+RMS,256ms 32ms overlap continuous on steady state signal TD feature set,150ms 50ms overlap continuous on steady state signal

Classier Channels /Classes ANN2 channel 4 MLP class classier Adaptive 2 channel 4 FUZZY class classier LDA 4 channel 6 class classier

Classier Delay Error 16-2% 250ms

101.3% 7.91.1%

<300ms

<300ms

GMM

4 channel 6 class classier 12 channel 11 class classier

6.71.1%

<300ms

LDA

96.3elbow 86.9wrist

elbow 220ms, wrist 380ms

Table 1: Summary of Myoelectric classier performances

1.1

EMG pattern recognition process

The pattern recognition process can be summarized to 5 steps as in gure 1. In implementing a classier it should be decided upon various key parameters and methods employed in each of the steps. The gure illustrates few of the key decisions necessary at each stage.

Figure 1: EMG pattern recognition process

LDA based EMG classication

While classication of EMG can be performed in various techniques. Literature suggest that Linear discriminant analysis based classiers performance is comparable to much stronger but rather expensive neural network based techniques. So for this project initial studies implements a LDA based classier to a myoelectric dataset.

2.1

Experiment

The work replicates main procedure reported in [2] with some minor additions. 4 surface electrodes placed on wrist exors, extensor, unlnar and radial deviators. These were used to capture resulting myoelectric signals during wrist exion, wrist extension, ulnar deviation, radial deviation and grip tasks. These tasks were performed sequentially and represented by indexes 0 to 5 in order. This one cycle of 5 tasks was repeated 20 times. The data sets were sampled at 1 KHz using an EMG recording system at the School of Human Kinetics and Recreation- Human movement analysis lab under expert supervision.

Channel 1

2 0 2 0 2 0 2 0 5 10 15 20 25 2 0 2 0 5 10 15 20 25 2 0 2 0 5 10 Time(s) 15 20 25

10

15

20

25

Channel 4

Channel 3

Channel 2

Figure 2: Raw EMG signal of 4 channels for 1 trial

2.2

Data Segmentation

The continuous windowing scheme proposed by Englehart et al. was utilized in the study. A continuous windowing exploits the full process capability of the available signal 3

processor which produces denser decision streams. The window length is kept at 32 ms and the process time delay used for advancing the window is set to 16 ms. These values were the optimal parameters reported in [2]. Although the process time delay should be set to the actual hardware speed for best results when used in embedded setting.

2.3

Feature Extraction

4 features were extracted from each EMG channel according to the data segmentation scheme selected. This includes Mean Absolute value MAV, Zero crossings (ZC), Sign changes (SC) and Signal length. This constitutes a 16 dimension feature vector for the pattern recognition task . The output vector carried an index representing the state (task) the arm was performing.

Ch.1 MAV

0.5

0 0 5 10 15 20 25 1 0.5 0 0 5 10 15 20 25 1 0.5 0 0 5 10 15 20 25 1 0.5 0 0 5 5 10 15 20 25

State 0 0 5 10 Time(s) 15 20 25

Figure 3: Feature 1 extracted from the 4 channels and corresponding state(task performed)

2.4

Ch.4 MAV

Ch.3 MAV

Ch.2 MAV

Class Extraction

Although there was a labeled dataset with an assigned class per 5 second window, the task was initiated from rest and terminated to rest during the 5 second period. By using threshholds at the corresponding muscles EMG level the period which the arm was resting 4

can be identied. But because of muscle fatigue and the stochastic nature of the signal corresponding to grip task, some class extractions would be erroneous. This could have been eliminated during experiment if the task was continually performed throughout the 5 second period.

2.5

Feature reduction

Linear discriminant analysis (LDA) was performed to reduce the feature space in to a simpler space with classiable data points. LDA reduces the 16 dimensional space in to 5 dimensions, in the 6 class classication case. As example gure 4 illustrates the reduced feature space for the 16 feature 3 class problem.
3
Rest S1

2.5

Wrist Flexion S2 Wrist Extension S3

1.5

0.5

0 0.3

0.2

0.1

0.1

0.2

0.3

0.4

Figure 4: Reduced feature space for three classes using Linear Discriminant Analysis For this particular study a Pairwise Linear Discriminant analysis was used. This increases the 16 feature 6 class classication problem in to 15 problems of bi-classication. At the same time reducing the feature space to 1 dimension. A rather interesting result was identied by analyzing the reduced dimensional space. All class given probabilities were resembling very closely to gaussian distributions except class 0 which was rest. This resembled a Gamma distribution. This observation gives the ability to appreciate Generative Learning algorithms as classiers to this data. Although logistic regression was used for classication in this study.

200 180 160 140 120 100 80 60 40 20 0 0.3 200 180 160 140 120 100 80 60 40 20 0 0.3 250 0.2 0.1 0 0.1 0.2 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5

200

150

100

50

0 0.4

0.2

0.2

0.4

0.6

Figure 5: Reduced feature spaces from pairwise linear discrimination, for class pairs 2-3 ,2-5 and 5-6 in order,with the class given probability distributions and logistic regression results for each pair.

2.6

Supervised Learning

A pairwise Logistic regression was performed with maximum likelihood parameter estimation. A gradient accent algorithm was used for Likelihood maximization(learning). Cost functions were observed, to verify learning and convergence. For supervised learning the rst ten trials of data was used. Figure 5 illustrates the results for 3 class pairs. 15 such discriminative functions were trained for this particular study. It is also possible to use other classiers such as neural networks, support vector machines etc. in the reduced feature space, for classication of data.

2.7

Validation

After learning the logistic function for each class, the last 10 trials of data were presented to this classier for validation. The classier calculates posteriori probabilities of each feature set to select the class with maximum posteriori probability of the discriminant function pairs. For human use it is acceptable to have a maximum delay of 300ms. But for processing of a decision 16ms is used. So maximum voting post processing scheme was used in a 128ms window (i.e 8 decisions). An 89% classication accuracy was observed using this method ,although higher accuracies is expected after the errors in class extraction is exploited (Figure 7). Even though optimal parameters for data segmentation reported in [2] was used in the study, it was observed that 64 ms data window improved the accuracy upto 92%. Further there was only little improvement observed between the disjoint and continuous windowing schemes.

2.8

Improvements

The errors present in class extraction deteriorates the classiers performance. So it is expected to perform the experiment such that, a task is performed throughout the predened time period to eliminate the requirement of class extraction. When the method is applied for amputees , only isometric muscle contractions will be present. So it is more suitable to use data sets with isometric muscle contractions for classier performance comparison. The classier used here employed pairwise discrimination, which enables visualization of the learning. But it would be ideal to implement the classication on all classes at once using softmax regression for speed of convergence.

2.9

Future work

As future work other classier performances will be tested against the LDA based method. Specically neural network based and support vector machine based classiers will be utilized. This can be further improved by using higher order feature sets such as auto regressive features and wavelet transformations. It is expected to improve the method for realtime classication taking in to account classier speeds and embedded myoelectric signal processors.

Before MV postprocessing(86.80%)

After MV postprocessing(89.76%) Actual state Classifier output

State

State 0 1000 2000 Sample 3000

100

200 Sample

300

400

Figure 6: Results of the classier output for the Validation dataset. 32 ms data window with continuous windowing scheme.(only 2 out of 10 cycles are shown)

References
[1] B. Hudgins, P. Parker, and R. Scott, A new strategy for multifunction myoelectric control, IEEE Transactions on Biomedical Engineering, vol. 40, pp. 8294, Jan. 1993. [2] K. Englehart and B. Hudgins, A robust, real-time control scheme for multifunction myoelectric control, Biomedical Engineering, IEEE Transactions on, vol. 50, no. 7, pp. 848854, 2003. [3] T. A. Kuiken, G. Li, B. A. Lock, R. D. Lipschutz, L. A. Miller, K. A. Stubbleeld, and K. B. Englehart, Targeted muscle reinnervation for real-time myoelectric control of multifunction articial arms, Jama, vol. 301, no. 6, p. 619, 2009. [4] F. H. Chan, Y. S. Yang, F. K. Lam, Y. T. Zhang, and P. A. Parker, Fuzzy EMG classication for prosthesis control, Rehabilitation Engineering, IEEE Transactions on, vol. 8, no. 3, p. 305311, 2002. [5] K. Englehart, B. Hudgin, and P. A. Parker, A wavelet-based continuous classication scheme for multifunction myoelectric control, Biomedical Engineering, IEEE Transactions on, vol. 48, no. 3, p. 302311, 2002.

Before MV postprocessing(90.8%) 6 6

After MV postprocessing(91.4%) Actual state Classifier output

State

State 0 1000 2000 Sample 3000

0 0 100 200 Sample 300 400

Figure 7: Results of the classier output for the Validation dataset. 64 ms data window with continuous windowing scheme.(only 2 out of 10 cycles are shown) [6] Y. Huang, K. B. Englehart, B. Hudgins, and A. D. Chan, A gaussian mixture model based classication scheme for myoelectric control of powered upper limb prostheses, Biomedical Engineering, IEEE Transactions on, vol. 52, no. 11, p. 18011811, 2005.