Derived Power Streams for Fault Detection and Condition-Based Maintenance

Nonintrusive electrical monitoring enables fault detection and diagnostics for power systems from an aggregate monitoring point. Faulty electromechanical equipment often produces unrecognizable electrical signatures that confound data-driven techniques. Fault signatures are typically unknown during training. For some loads, analytical models and simulation can predict fault signatures, facilitating their inclusion in the training set. However, this approach scales unfavorably with the number of parameters associated with faults. An alternate outlook recognizes that faults often change the power system conditions and the quality of power delivered. Assumptions about the character and quality of the local grid should be called into question during faulty load behavior. The objectives of this work are to correlate these changing grid behaviors with load and fault signatures as a step towards automatic fault detection. Our methodology involves creating supplemental “derived” data streams distilled from nonintrusive electrical measurements that can correlate changes in grid conditions and quality with load operation. Using real fault data from shipboard microgrids, our key findings illustrate that combining these derived streams with nonintrusive power data enables both nonintrusive load identification and fault diagnosis. Finally, this work presents a vision for implementing an integration of these streams with physics-based fault models in an automatic fault detection system.