A common representation for phase-resolved pattern is
based on the phase-window method in [12], [15], [11], [17].
The phase-window method divides the power cycle with 360◦
into several small phase windows and then generates some
features for each phase window. For example, we have 360
windows if each phase window has a size of 1◦. Assume
that we have three features for each phase window, namely
the number of pulses, maximal amplitude and the average
amplitude in this phase window. We can thus represent the
phase-resolved pattern with a 360 × 3 = 1080 dimension
vector. Note that the feature vector for either a waveform
(1000 dimension in Figure 2(A)) or a phase-resolved pattern
(1080 dimension as introduced above) has high dimensions.
Due to the curse of dimensionality, it is difficult to analyse,
understand and visualise such high dimensional data. Feature
extraction techniques for dimensionality reduction would thus
be important before we can analyse the PD data [9].