We apply unsupervised machine learning to three years of continuous seismic data to unravel the evolution of seismic wavefield properties in the period of the 2009 L’Aquila earthquake. To obtain sensible representations of wavefield properties variations, wavefield features, i.e. entropy, coherency, eigenvalue variance, and first eigenvalue, are extracted from the covariance matrix analysis of continuous array seismic data. The defined wavefield features are insensitive to site-dependent local noise, and can inform the spatiotemporal properties of seismic waves generated by sources inside the array. We perform a sensitivity analysis of these wavefield features and build unsupervised learning based on the uncorrelated features to track the evolution of source properties. By clustering the wavefield features, our unsupervised analysis avoids explicit physical modeling (e.g. location of events, magnitude estimation) and can naturally separate peculiar patterns solely from continuous seismic data. The unsupervised learning of wavefield features reveals distinct clusters well correlated with different periods of the seismic cycle.