SUMMARY

Nowadays cheap computational and sensing resources availability for real world systems has made such complex systems monitoring, based on large scale system’s high dimensional acquired data, more accurate and useful for applications of system modeling, prediction, and diagnostics via learning based function approximation. While the hardware resources and data production are progressing at an enormous rate, there is an ever increasing need for autonomous, adaptive, and intelligent decision making in engineering and other systems.

The objective of this thesis will be to develop algorithms and methods for learning based model identification for prediction and diagnostics applications.

First, we will explore deep Convolutional Neural Network (CNN), and will propose an improvement to its gradient vanishing problem. Second, we will present a novel deep network architecture that will be composed of multiple units of two-hidden-layer based Multilayer Perceptron (MLP) neural network that will learn feature extraction followed by feature to output mapping. Thirdly, we will present a novel data transform based approach to develop a method of learning the system model. These approaches are aimed to alleviate problems and limitations in one of the best state of the art methods of model learning. Finally, we will apply these methods on different data types for the prediction and diagnostics problems in robotics and other domains.