**3. Categorization of EMG based control methods**

Control methods of assistive robots based on EMG can be categorized mainly according to the input information to the controller, architecture of the control algorithm, output of the controller and other ways. Figure 3 shows the ways of categorization of EMG based control methods. In this chapter, the EMG based control methods of assistive robots are categorized based on the architecture of the control method. Considering the EMG signal processing method in the architecture of the controller the EMG based control methods can be categorized mainly as pattern recognition based and non-pattern recognition based [5, 25]. Control methods of many assistive robots are designed with pattern recognition based control methods and it provides an accurate control action compared to non-pattern recognition based EMG based control [5]. However, several intermediate steps (see Fig.4) are applied with a pattern recognition method such as data segmentation, feature extraction, and classification [5, 25]. The accuracy of pattern recognition based control is greatly improved by methods of feature extraction and classification [5, 25]. Robots such as, SUEFUL-7 [6], NEUROExos [27], W-EXOS [11], DEKA Arm [21], Saga Prosthetic Arm [22], Manus Hand [23] are on pattern recognition based control. The non-pattern recognition based control method involves only a few steps (see Fig.5).

'Amplification' is wrong.

Please replace Fig.6 from given figure below. In the previous figure the spelling for the

**Figure 6.** Steps of pattern recognition based EMG processing

EMG processing in pattern recognition based control [5] and non-pattern recognition based control are further illustrated in the next subsections.

3

**3.1. Pattern recognition technique**

recognition based EMG based control method.

hence better for the data segmentation [5].

The different steps coming under pattern recognition such as data acquisition, data segmen‐ tation, feature extraction and classification can be further broken down into sub areas consid‐ ering available options as shown in Fig. 6. Next subsections present further categorization of data segmentation, feature selection and classification respectively coming under the pattern

Recent Trends in EMG-Based Control Methods for Assistive Robots

http://dx.doi.org/10.5772/56174

243

**Data segmentation**: The EMG signal has two states: transient and steady. In the transient state, the muscle goes from rest to a voluntary contraction level [5]. Constant contraction of the muscle can be seen under the steady state. In addition the EMG signal in the transient state shows a large deviation of error compared to the steady state level. Therefore, in many cases, the steady state signal is used for the analysis of EMG. For the better result of data segmenta‐ tion, the selected time slot should be equal to or less than 300ms. This includes segment length and processing time to generate the control command. In addition bias and variance of features can be minimized by selecting adequately a large time slot and it contributes to better classi‐ fication performance [5, 12, 19]. Data segmentation is carried out with two major techniques: overlapping segmentation [5, 25], and disjoint segmentation which uses segments with predetermine length for feature extraction. Also processing time is a small portion of segment length and thus processor is idle for remaining time of the segment. The new segment slides over the current segment and has small incremental time for overlapping segmentation technique. According to [25], overlapping segment method increases processing time and

**Feature selection**: Due to the large number of inputs and randomness of the signal, it is impractical to feed the EMG signal directly to the classifier [6, 18]. Therefore, it is necessary to create the feature vector, where sequence is mapped into a smaller dimension vector. Success of any pattern recognition problem depends almost entirely on the selection and extraction of features. According to the literature, features fall into one of three categories: time domain, frequency domain and time scale domain [25]. Many assistive robots use time domain analysis for feature extraction and in most cases, RMS calculation is adapted for feature extraction [6, 35]. Assistive robots based on frequency domain and time frequency domain were scarce.

**Classification**: Extracted features need to be classified into distinctive classes for the recogni‐ tion of the desired motion pattern [5]. Several external factors, such as fatigue, electrode position, perspiration and posture of the limb may causes changes in the EMG pattern over time. Therefore, this leads to large variations in the value of a particular feature. The important feature of the classifier is the ability to identify the unique feature throughout the varying pattern due to other influences. The speed of the classifier is an important aspect for generating required output from the controller. Further, training of the classifier is another way to improve the response of the control system of the assistive robot. Depending on the performance of the subject, practice required for rehabilitation etc can be customized through a training of the classifier and it produces the expected rehabilitation training for the patient too. According to the literature, several methods are used for EMG classifications. Some of them are Neural network [5, 35], Fuzzy logic [25], Neuro-fuzzy logic [5, 25], Probabilistic approach etc.
