Single-molecule nanocircuits based on field-effect transistors (smFETs) are emerging and promising nano-bioelectronic sensors for the functional detection of molecular dynamics involved in biochemical transformations, in particular for applications in cancer thanks to a potentially better understanding of some hidden and complex molecular interactions. In fact, functionalized carbon nanotubes have been recently exploited to probe molecular events occurring at a single molecule scale with ultra high sensitivity and specificity, such as nucleic acids hybridization, enzyme folding in catalysis reactions, or protein-nucleic acids interactions. Extracting the kinetics and thermodynamics from such single-molecule dynamics implies robust analytic tools that can handle the complexity of the sensed reaction system changing between transient and steady-state molecular conformations, but also some challenging signal specificities, such as the multi-source composition of the recorded signals, the mixed and high-level noises, and the sensor baseline drift, leading to non-stationary time series. We present a new smFET data analysis framework, based on a compressive feature learning scheme to optimize unsupervised idealization of smFET traces, by a precise and accurate molecular events detection and states characterization algorithm, tailored for non-stationary signals at high sampling rate and long acquisition periods, without any prior knowledge on the data generating process nor signal prefiltering. Experimental results show the accuracy and robustness of our trace idealization algorithm to stochastic state-space models, and better performances than commonly used hidden Markov models.