Electrical Monitoring of Electromechanical Systems

Abstract

Electromechanical systems provide the world’s backbone for generating and using energy. Electromechanical systems can also experience an innumerable set of failures, causing induced wear and wasted energy, or eventually a complete failure of a critical piece of equipment or system. Degradation or other faults are often associated with subtle but observable changes in electrical consumption. A nonintrusive load monitor (NILM) is a convenient tool for electrical monitoring, in which all loads connected downstream of an electrical panel are monitored with a single set of current and voltage sensors. If collated in a useful way, nonintrusive electrical data can make diagnostic information more easily attainable and improve the efficient operation of critical machines.

Ensuring correct nonintrusive identification of load operation is a challenge in varying operating conditions and fault scenarios. Most nonintrusive load monitoring research assumes that data is static over time. Also, ground truth labels are a scarce resource in industrial scenarios. Thus, a pattern classifier must train on a limited dataset not representative of long-term operation. This thesis employs an understanding of the physics and time-dependency behind changing load behavior to inform pattern classification. New statistical feature extraction techniques are presented for loads with time-varying operation. Results are demonstrated with laboratory experiments and case-studies from NILM installations onboard various marine microgrids.