Classifier economics of Semi-Intrusive Load Monitoring

Non-Intrusive Load Monitoring (NILM) and Semi-Intrusive Load Monitoring (SILM) are fast developing techniques for devices operation recognition in system monitoring. Many traditional researches focus on feature space improvements for better recognition accuracy and classifier/meter quantity reduction. But practically, cost of each classifier/meter will influence the optimal NILM/SILM solution. A feature space with better accuracy in NILM may require more cost than a SILM solution with multiple classifiers with simpler feature spaces. Facing this issue, this paper initiates a new classifier network construction method for NILM/SILM. Instead of creating a classifier for NILM or SILM, this method helps decision maker to select different types of classifiers and optimally allocates the classifiers' positions. In this method, economics of each type of classifier is considered to ensure decision maker's cost reduction. A combinatorial optimization problem is established on a tree-type model to the optimized classifier network. Numerical studies on a public data set REDD and an industrial operational data are implemented to support the feasibility of the method.