• No results found

Algorithms in Upper Limb sEMG Based Sensing and Adaption

When considering pattern recognition based approaches towards hand motion recog- nition there exists two major branches of sensing algorithms based on those that use feature based information and those that use more modern deep learning based ap- proaches. A particular focus here shall be placed on sEMG based sensing.

2.3.1

Hand Crafted Feature Based

For feature based sensing there is a long history of different classification strategies and implementations which have been widely studied towards bio signal based sensing in offline and online environments. Typically many researchers will opt for Linear Dis- criminant Analysis (LDA) or Support Vector Machines (SVM) based classifiers likely due to their reliability, ease of implementation, and likely due to their prominence in research [76, 77, 78, 79]. Although LDA and SVM are popular methods they may not promise the best quality results. Within the state of the art other frequently used classifiers include Quadratic discriminant Analysis (QDA)[80], K-Nearest Neighbors (KNN)[81], Multi-layer Perceptron Networks (MLP)[82, 83], Hidden Markov Mod- els (HMM)[84], Artificial Neural Networks (ANN)[85], Fuzzy Logic (FL), Gaussian Mixture Models (GMM) [86].

All of the above algorithms have demonstrated excellent performance in labora- tory conditions for sEMG based hand motion recognition for various forms of hand motions and object manipulation, often seeing accuracy scores of over 90%. The real challenge with these sensing approaches exist when moving towards clinical environ- ments and longer term use. As described previously, there exists many variable traits

of sEMG signals from changing physiological aspects of the user and environmental factors during sensing. Towards clinical applications, researchers have observed the performance of varying algorithms when considering fatigue [87, 88]. Whereas other researchers have explored the capabilities of various algorithms towards inter day use and others for the feasibility of inter subject use.

A common challenge seen in these tests is that of how to best approach the seem- ing randomness of bio signals. One solution that has been experimented with is that of constructing datasets which have larger training datasets that may contain these variable changes. An alternative approach is that of utilizing adaptive algorithms to update a trained model to whichever changes are seen. A promising route is that of semi adaptive algorithms which require user input to update a model with new exem- plars however a limitation of these models is that intervention is required which may cause frustration during use. Unsupervised adaptive methods have also been proposed to change during active use. A further investigation into the challenges of long term sEMG based sensing algorithms is performed in chapter 3.

2.3.2

Deep learning

While hand crafted feature based approaches have long been the dominant method for sEMG based pattern recognition, recent years have seen a growing interest in deep learning. Typically, deep learning approaches had been connected to computer vision, particularly that of object detection. The process of deep learning based approaches to progressively extract low level input and convert it into high level features is particu- larly valued in computer vision. It is this feature learning trait of deep learning trait of deep learning that holds particular promise over traditional hand crafted features for sEMG based hand motion recognition. Although deep learning methods are capable of extracting high level features from a dataset, the nature of deep learning networks demands a large dataset otherwise they risk overfitting and lack of generalisation. The risk of overfitting is a particular challenge when converting deep learning approaches from computer vision to sEMG based pattern recognition due to the relatively small datasets traditionally used in sEMG sensing. While it is possible to construct a large enough sEMG dataset for deep learning, large datasets such as the Ninapro dataset [89]

provide both large volumes of sEMG data and a wide array of gestures for benchmark- ing methods.

Deep learning approaches are often divided into three categories of unsupervised pre-trained networks (UPNs) such as Deep Belief Networks, convolutional neural net- works (CNN), and recurrent neural networks (RNN). Shim et al. [90] utilized a Deep Belief Network (DBF) on locally collected two channel data with 7 grasps, demon- strating superior performance of a DBF against LDA and SVM. Of the forms of deep learning network, convolutional neural networks appear to attract the most interest in sEMG based sensing and other fields. It was in 2016 when CNN’s begun to be ap- plied to sEMG based sensing, where a CNN was demonstrated to achieve higher inter- subject motion recognition than SVM [91]. Geng et al. [92] further applied CNN’s on single sEMG images from a frame of data to enable instant recognition, which was further improved across a 40 frame window with majority vote. Inter-day evaluations of CNN’s further show promise as demonstrated by Du et al. [93] on two day interday data compared against feature based classifiers. The long term viability for interday classification with CNN’s was demonstrated by Rehman et al. [94] across a period of 15 days. Rehman et all posed that CNN approaches may overcome the feature calibra- tion challenges of hand crafted feature based approaches.

While Deep learning based approaches demonstrate promise over traditional meth- ods, the resource intensive nature of deep learning may pose challenges in clinical environments, such as during user training.

With any combination of sensing modalities and algorithms to infer user intent, a particularly vital aspect in sensing and adaption is that of user training and rehabili- tation. It is important to verify that a training approach can ensure repeatability and applicability to a users daily life. Furthermore, an ideal prosthesis would provide in- tuitive control alongside feedback that can best guide its user. The next section will firstly focus on present metrics to gauge rehabilitative performance and to train users. Secondly, the next section will include an overview and evaluation of prosthesis haptic feedback methods.