Abstract
Optimal sensor placement is a challenge in many engineering design applications, especially within the field of prognostics and health management. Recently, data-driven approaches have become a staple for solving and addressing these challenges. Machine learning techniques have been applied to solve complex optimization problems in the field of signals processing. However, these methods require a substantial amount of data, which can be difficult to obtain. In addition, the design space may be extremely large, so a deterministic approach may not be possible. Therefore, there is a need for probabilistic frameworks that can simultaneously train a classifier for detection of faults as well as selecting new designs for optimal placement. In this paper, the proposed methodology contains a genetic algorithm embedded with a clustering algorithm to simultaneously train the classifier and determine a sensor network. This novel structure is implemented for detecting short-winding faults of a permanent magnet synchronous motor using magnetic field sensors. The training data is simulated using a finite element model, and the design space is extremely large. Nonetheless, the results of the proposed methodology show accuracy for detection of faults.