Increased demand on power transformers, due to the extension of equipment lifetimes and a trend to lower costs, has led to a need for condition based maintenance. However, there are currently no established techniques to accurately monitor and diagnose faults in real-time while the transformer is on-line. A major factor in the degradation of transformer insulation is partial discharging which, if left unattended, will eventually cause complete insulation failure. A partial discharge detection system (PDDS) is presented for electrically identifying partial discharge pulses and determining the number of partial discharge sources while the transformer is on-line and subjected to external interference. Current and voltage transducers are placed on the transformer bushings rendering it unnecessary for the transformer to be disconnected or opened. Partial discharge pulses are identified within narrowband and pulse interference using digital signal processing techniques which include an automated Fourier domain threshold filter and a continuous wavelet directional coupling filter. If multiple sources are active, the partial discharge pulses are grouped by a clustering neural network. Two 7.5kVA, 11kV transformers were used to conduct testing of the PDDS. High recognition rates of artificial partial discharge pulses were achieved in an off-line transformer, while a previously unknown natural partial discharge source was identified in the second energised transformer.