This work investigates mathematical and computational methods suitable for analysing data emanating from large cyber physical systems. Embedding the governing equations for the system behaviour, especially dynamics, ensures analysis solutions which are consistent with the physics of the system. The developed methods also deal with the implicit uncertainty fundamentally associated with perturbed data. Symbolic data analysis is investigated as a means of establishing a consistent computational approach to perform automatic unsupervised identification of structures in multi-channel time series data. This is achieved by mimicking techniques from the evolution of natural language. The validity of the approach is demonstrated in an application to automatic operations recognition. Particularly interesting in this context is the identification of human interaction with the system via structure embedded in the data. Additionally, this thesis considers the issue of characterizing sensors and quantifying their behaviour, in particular modelling their uncertainty. This is fundamental since errors entering via the interpretation of sensor data will propagate through the entire analysis cycle. The established methods and techniques are integrated into a framework to support end-to-end applications, i.e. from the data acquisition to the presentation of the results. A software tool, developed within this work, extends the framework to support the data analyst in the handling, analysis and visualization of large multi-dimensional time series together with the computational results. The conducted research is presented as a collection of papers woven together with introductory texts and some extensions to form a complete thesis.