Fit first, ask questions later: Interpretable Unsupervised Learning for High-Volume, High-Dimensional Data in Engineering.
Abstract
In modern engineering practice we are often faced with large volumes of high-dimensional data. Typical settings where such data are encountered, are outputs and/or inputs of Monte-carlo simulations, physical experiments with refined measurements and raw condition monitoring data streams. Variation in data, regardless of the engineering setting involved, posses three fundamental questions for analysis: “What” varies in the data, “How/How much” it varies, (identification and quantification of factors of variation), and finally “Why” it varies (interpretation). In traditional engineering modeling settings, in order to simplify the process from data to insight and models, we are forced early on to form restrictive hypotheses for all of these questions by explicitly modelling physical phenomena or by targeting small subsets of the data. In effect, we are potentially losing valuable insight due to the reduction process. Unsupervised learning is the family of techniques for building models directly from data, which discover the “what” and “how”, allowing us to interrogate them in order to discover the “why”. In this talk, it is going to be demonstrated how unsupervised learning can yield engineering relevant insights. The techniques discussed are Pricnipal Components Analysis and the emerging deep neural network based Variational Autoencoders. The applications presented are going to be on fatigue damage accumulation in wind turbine blade spar-caps and on the effects of downstream vortices (wakes) in windfarms by considering simulated farm monitoring data.
