This thesis investigates selection of time domain (TD) signal features for myoelectric signal (MES) based control of motorised hand and wrist prostheses. A signal feature represents a distinguishing property of a MES to be used in pattern recognition algorithms. In particular, TD features reflect the mathematical functions and physical expression of the transient signal waveform with respect to time. Extracted features capture the structural details of a MES, minimise loss of information upon conversion, and simplify movement classification. The advantage of TD features is that they produce lower dimensional input vectors while maintaining sufficient accuracy of various movements if adequate information is provided. Feature sets as a solution to gather information in MES based control has not been thoroughly studied in the literature. We aim to develop methods to elevate the use of TD features and suggest a comprehensive feature set that is helpful in pattern recognition. Myolectric signals used in this study were from the BioPatRec database, an open source platform for research and control of artificial limbs via pattern recognition using bioelectric signals. This database is named as 10mov4chUntargetedForearm comprising data on10 hand and wrist movements acquired by 4 bipolar sEMG channels from the left or right forearm. Based on feature selection (FS) which preserves information of the MES, we propose three methods, namely a genetic algorithm (GA), class relevant criteria and a self-organising feature map (SOFM) to assemble feature sets from TD features of twenty one candidates. To evaluate these feature sets, we implemented three pattern recognition algorithms, particularly the Multilayer Perceptron (MLP), Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) algorithms. The reported movement accuracy and Wilcoxon p value demonstrated that the proposed feature sets consistently outperformed a typical feature set found in the literature; in particular, improved the accuracy of poor quality datasets from 85% to 93%. The thesis has made a thorough investigation of TD features contributing in three categories. Firstly, we developed a variety of independent methods for FS. It was noticed that FS has been limited in meta-heuristic searches in the literature. We have demonstrated that there are several solutions that use potential TD features to assemble a feature set to be used in pattern recognition. Secondly, we have shown that statistical tests can be successfully applied in FS. Thirdly, we explored an investigation of data along time series vectors instead of analysing it conventionally by time segmentation. The success of this method suggests a new way that may value further inspection. In brief, this thesis presents possible solutions for TD feature based pre-processing of the input of pattern recognition algorithms for prosthetic control. It provides immediate accuracy improvement through a replacement of feature sets and further implementation in methodology for FS.