Frequency Invariant Transformation of Periodic Signals (FIT-PS) for Classification in NILM

• Published in: IEEE transactions on smart grid Vol. 10; no. 5; pp. 5556 - 5563
• Main Authors: Held, Pirmin, Mauch, Steffen, Saleh, Alaa, Ould Abdeslam, Djaffar, Benyoucef, Dirk
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.09.2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE); Institute of Electrical and Electronics Engineers
• Subjects: Computer Science; Current measurement; Electric potential; Feature extraction; feedforward neural net; Harmonic analysis; Invariants; LSTM net; non-intrusive load monitoring; Representations; Signal and Image Processing; Signal classification; signal form; Signal representation; Steady-state; supervised classification; Switches; Transformations; Transient analysis; Variations; Voltage; Voltage measurement; Feature extraction; Non-intrusive load monitoring; Feed forward neural net; Signal form; Supervised classification; LSTM net
• ISSN: 1949-3053; 1949-3061; 1949-3061
• DOI: 10.1109/TSG.2018.2886849

This paper presents a new signal representation called frequency invariant transformation of periodic signals (FIT-PSs) in the context of non-intrusive load monitoring (NILM). Compared to former approaches where a conglomeration of different signal forms are used, the presented approach is based on a single signal form containing all information. The core idea of this paper is to use the original current waveform relative to the reference voltage as a signature for NILM. In general, the relation of sampling and grid frequency is subject to continuous fluctuations. Therefore, FIT-PS converts uncorrelated sample data to a fixed multiple of the grid frequency. The advantages are that the information of the current signal, as well as the phase shift between voltage and current signal, are completely contained in the FIT-PS signal representation. For classification, a neural net was applied to the home equipment laboratory dataset 1. Features created by FIT-PS are superior to the standard features. With 18 different devices, a detection rate of up to 90% is achieved. In particular, when several consumers are active at the same time, the new signal representation is much more robust and leads to a better detection rate. However, a long short-term memory net with FIT-PS signal representation provides the best results.