NEURAL NETWORKS GOT OUT ON A LIMB

Currently available prosthetic hands are very primitive. They are usually limited to movement in a single plane to produce a pincer-like action in which the thumb and coupled ring and middle finger touch together in a three-jaw chuck configuration. These devices tend to use myoelectric signals (MES) – electrical potentials detected on the skin surface that are generated by underlying muscle activity. These signals are detected from the upper arm as a switch, to open and close the hand.

Now, research conducted at the University of Plymouth could allow more sophisticated prosthetic devices to be produced. Designers there have concentrated on studying the nature of the MES generated by muscles in the upper and lower arm and the correlation of the MES to a range of different hand movements. The research has revealed that, for a given muscle site, both the amplitude and frequency content of the MES varied for different hand movements. However, the MES signature for a particular movement is reasonably consistent and repeatable.

The challenge was to establish whether an MES from one or more muscle sites could be taken and ‘decoded’ so that the prosthetic arm servo drive could respond with an appropriate hand movement.

To help out, the developers are using artificial neural networks (ANNs) systems. The fundamental feature of any neural network is that it is composed of a large number of interconnected processing units, arranged as input, hidden and output layers. Each interconnection has an associated ‘weight’, whose value is determined during the training phase. Here, the outputs from the network are compared with some desired values and the weights adjusted using back propagation to minimise errors.

In the case of an artificial hand, the output from the network is a code that represents a hand movement. The input is a processed form of the MES signature. This processing was undertaken by splitting the frequency content of the MES into a set of discrete frequency values using either an FFT approach or the application of digital band-pass filters. Either technique produces a set of numbers that can be used to initially train the network for given desired hand movements. Once trained, the network recognises MES frequency signatures as representing certain hand motions.

A prototype network has been designed and trained on a member of the research team. With this system, a computer simulation of a hand is connected to the output of the network and represented on a monitor. In this way, movements of the researcher’s hand are replicated by the simulated hand on the computer screen.

Currently, an investigation is underway to map the total surface of the upper and lower arms to identify optimum MES sites. In addition, tests are being conducted on a range of males and females of varying ages to discover how amplitude and frequency of the MES is a function of the muscular activity of different human beings. Additionally, the software side of the project is being adapted to fit within a digital signal processing chip.

University of Plymouth

Tel: Plymouth (01752) 232637