Nowadays, the term machine learning gathers together a set of methods that can automatically detect patterns in data, and then use the uncovered patterns to predict future data, or to perform other kinds of decision making under uncertainty [118]. Controlling the mobility of the human body with the aid of articial limbs is a complex goal as the parameters to be controlled changes very often. These parameters include variable body weight, unpredictable height gain, change in body structure and reex action based movements. Determination of Cartesian space and joint space in powered prosthetic limbs is of great importance and there is a need to improve the stability of the mechanism. Non-intuitive control is one of the few reasons amputees have lost intrest on the use of their prostheses [30]. Switched or gated control is considered to be slow and non-intuitive, requiring both time and sustained cognitive eort on the part of the user [119]. Challenges and limitations of conventional or traditional control methods have motivated researchers to think in other terms, including the implementation of intelligent systems.
Extensive work done so far demonstrated that an adaptive neural control scheme is particu- larly suitable for controlling highly uncertain, nonlinear, and complex systems [120]. The control scheme can learn the full dynamics of a non-holonic system on-line, and can deal with parametric as well as non-parametric uncertainties, yet guarantees tracking errors asymptotically converge to zero [121]. Miao et al. [122] developed a recurrent neural network control system without requiring explicit knowledge of the system dynamics and the results were convincing for adaptive control for lower limb control.
Furthermore, [123] suggested that predictions are a key component of intelligence and neces- sary for accurate motor control. In reinforcement learning, such predictions can be made through general value functions (GVFs). These are temporally extended predictions about a signal of interest that have been applied to building up real-time anticipatory knowledge in relation to human-machine interactions. Their main contribution was on improving robotic articial limb performance through real-time learning and utilisation of temporally extended predictions. This was achieved through the use of multilayer predictions, that is, predictions based on predictions using simulations.
According to Degris et al. [124] reinforcement learning methods are often considered as a potential solution to enable a machine to adapt to changes in real time to an unpredictable environment. When developing a control architecture, one can choose to have hierarchies based on tasks, for example sitting, standing and walking. If well-structured, an o-line learning al- gorithm can be developed so as to switch between a specic set of decisions depending on the hierarchy being executed. A novel idea was proposed by [125] regarding 3D model retrieval and
classication by semi-supervised learning with content-based similarity. Using this technique, status of terrains can be identied and classied and a set of decisions can easily be activated. However, such a system has limitations when applied to sEMG signals.
The most common machine intelligence feature used in active limbs is Articial Neural Net- works (ANN). According to Elsa [126] ANN retains two characteristics of the brain as primary features: the ability to learn and to generalize from limited information. The approximation between the eld of ANNs and probability theory has led to a new domain called Statistical Machine Learning (SML) [127]. It is clear from these previous applications of Feed Forward Articial Neural Networks (FFANN) that once the network is trained using input−output ex- amples, it learns and generalises the input−output relation. As a result, it serves as a surrogate model to prevent repetition of time-consuming calculations. A trained neural network can make a real-time prediction of output for a new set of inputs or a real-time classication of a new input pattern [128].
Learning is undoubtedly one of the most relevant contributions of ANNs to the sphere of information processing systems. Three basic classes of learning can be considered as supervised, unsupervised and reinforcement and they allow for [129]:
Implementing a process with no need to know the mechanism model that lies beneath. Using the same model to deal with dierent tasks.
Adapting the system to changes in the surrounding environment. The basic tasks that can be performed through a learning process are [130]:
Pattern association. Pattern recognition.
Function approximation, including system identication, parameter estimation and inverse modelling.
Computation of the suitable parameters of a control system.
Morita et al. [131] managed to achieve torque control of each joint in the upper limb. The joint torque was estimated from many EMG signals using an articial neural network, hence the learning system was based on the feedback error schema. Apart from being o-line learning, all these systems learn from a policy already well known to the problem using pattern gener- ator type policies. However, there is a need to develop a new prediction method to overcome the problems of the training process in articial neural networks. Therefore, the requirements to develop models customised for several applications still exist regardless of the fact that the architecture of the ANN is universal.
The applicability of machine learning in rehabilitation systems is highly limited by hardware availability and the need for miniaturisation of the available technology [7]. Most of the afore- mentioned techniques were applied on oine systems using the Matlab software platform and large working stations. However, for clinical viability and wearable device technology size, power consumption and weight of hardware do matter most. In this study, machine learning will be used for pattern recognition in EMG signal classication. Signal classication, if well utilised, will assist in correctly predicting the partcipant's intentions in real-time, thereby improving active limb's adaptability and robustness [132].