Date of Completion


Embargo Period



data mining; global optimization; multi-objective optimization; machine learning; deep neural nets

Major Advisor

Jiong Tang

Associate Advisor

Horea Ilies

Associate Advisor

Chengyu Cao

Associate Advisor

Xu Chen

Associate Advisor

Ying Li

Field of Study

Mechanical Engineering


Doctor of Philosophy

Open Access

Open Access


The recently rapid advancements in sensing devices and computational power have caused paradigm shift in engineering analyses: data driven and sampling-based approaches play more and more important roles. For example, with sampling-based global optimization techniques, mechanical components can be refined and redesigned; system parameters can be accurately identified. With advancements in data mining and machine learning, underlying input and output relations of complex systems can be developed to make predictions with unseen data or to expedite analytical process. In this dissertation, advanced optimization and machine learning techniques are developed and employed to multi-facet engineering tasks from system design to identification.

First research task concerns a type of configuration designs, where the volume occupied components is to be minimized along with other objectives such as the length of connectivity lines. The formulation of computationally tractable optimization is difficult in practice as the objectives and constraints are usually complex. Moreover, the design optimization problems usually come with demanding constraints that are hard to satisfy, which results in the critical challenge of balancing feasibility with optimality. We develop an enhanced multi-objective simulated annealing approach, MOSA/R, to solve this problem. A versatile and efficient re-seed scheme that allows biased search while avoiding pre-mature convergence is designed in MOSA/R. Some generalization studies of the algorithm have also been carried out. In this second task, we exploit the impedance/admittance change of a piezoelectric transducer bonded to a host structure, aiming at the identification of system damage. To find a small set of solutions for such an under-determined system that indicates the true damage scenario, we cast the damage identification problem into a multi-objective optimization framework. With damage locations and severities as unknown variables, one objective function is the discrepancy between first-principal model predictions and actual measurements. The sparsity of the unknown variables is chosen as another objective function, deliberately, the l0 norm, because damage occurrence generally affects a small number of elements. A multi-objective algorithm (DIRECT) is devised to facilitate the inverse analysis where the sparsity is further emphasized by sigmoid transformation. As a deterministic technique, this approach yields repeatable and conclusive results. The third task concerns early diagnosis of gear transmission, which is challenging because gear faults occur primarily at microstructure but their effects can only be observed at a system level. The performance of a fault diagnosis system depends on the features extracted and the classifier subsequently applied. Fault-related features are conventionally identified based on domain expertise, which are system-specific. On the other hand, although deep neural networks enjoy adaptive feature extractions and inherent classifications, they require a substantial set of training data. We present a deep convolutional neural network-based transfer learning approach, which not only entertains preprocessing free adaptive feature extractions, but also requires only a small set of training data.