This thesis develops a robust methodology for on-line identification of power system dynamic signature based on incoming system responses from Phasor Measurement Units (PMUs) in Wide Area Measurement Systems (WAMS). Data mining techniques are used in the methodology to convert real-time monitoring data into transient stability information and the pattern of system dynamic behaviour in the event of instability.The future power system may operate closer to its stability limit in order to improve its efficiency and economic value. The changing types and patterns of load and generation are resulting in highly variable operating conditions. Corrective control and stabilisation is becoming a potentially viable option to enable safer system operation. In the meantime, the number of WAMS projects and PMUs is rising, which will significantly improve the system situational awareness. The combination of all these factors means that it is of vital importance to exploit a new and efficient Transient Stability Assessment (TSA) tool in order to use real-time PMU data to support decisions for corrective control actions. Data mining has been studied as the innovative solution and considered as promising.This work contributes to a number of areas of power systems stability research, specifically around the data driven approach for real-time emergency mode TSA. A review of past research on on-line TSA using PMU measurements and data mining is completed, from which the Decision Tree (DT) method is found to be the most suitable. This method is implemented on the test network. A DT model is trained and the sensitivity of its prediction accuracy is assessed according to a list of network uncertainties. Results showed that DT is a useful tool for on-line TSA for corrective control approach. Following the implementation, a generic probabilistic framework for the assessment of the prediction accuracy of data mining models is developed. This framework is independent of the data mining technique. It performs an exhaustive search of possible contingencies in the testing process and weighs the accuracies according to the realistic probability distribution of uncertain system factors, and provides the system operators with the confidence level of the decisions made under emergency conditions. After that, since the TSA for corrective control usually focuses on transient stability status without dealing with the generator grouping in the event of instability, a two-stage methodology is proposed to address this gap and to identify power system dynamic signature. In this methodology, traditional binary classification is used to identify transient stability in the first stage; Hierarchical Clustering is used to pre-define patterns of unstable dynamic behaviour; and different multiclass classification techniques are investigated to identify the patterns in the second stage. Finally, the effects of practical issues related to WAMS on the data mining methodologies are investigated. Five categories of issues are discussed, including measurement error, communication noise, wide area signal delays, missing measurements, and a limited number of PMUs.