Machine learning assisted system identification of a composite plate

Abstract

Parallel advancements in the fields of system identification and machine learn ing have been occurring for years now, but only recently have both the concerned communities realized how they been have trying to address fundamentally similar types of problems more often than not. Not only have almost all the major research fields started exhibiting a significant bias towards machine learning techniques for solving complex problems, but also a whole new market for machine learning based technology has emerged with the rise of powerful computer systems. To summarize the research work done in this thesis, an Artificial Neural Net work (ANN)-based approach has been proposed to identify the mechanical properties an orthotropic composite plate, modelled using Finite Element Analysis (FEA) and sampled using Latin Hypercube Sampling (LHS), in frequency as well as time do main. The two networks employed for the same are Multi-Layer Perceptron (MLP) and Radial Basis Function Network (RBFN), both of which predict, to a certain degree of error, the four parameters characteristic of any composite plate: Young’s moduli in two directions, Poisson’s ratio, and shear modulus. Eventually, frequency domain identification using the MLP network was accepted as a much superior model as compared to the RBFN with an accuracy of 2.7% and a training time of less than 7 minutes averaged over the said parameters. Keywords: System Identification; Finite Element Analysis (FEA); Latin Hypercube Sampling (LHS); Artificial Neural Network (ANN).