Comparison of recurrent neural network-based autoencoders for condition monitoring of individual devices from summarized electrical power signals

Abstract

Listen

Translate article

Traditionally, condition monitoring of manufacturing devices requires the installation of multiple submeters to obtain necessary data from individual devices, such as measuring electrical power demand. While effective, this process can be expensive and resource-intensive. To address the limitations of submetering, Non-intrusive load monitoring (NILM) is used to acquire the power demand data from the main electrical supply of a device group as a summarized power signal. The summarized signal is disaggregated to obtain the power signal of each device using deep learning algorithms such as Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Subsequently, the individual power signals can be used as an input to develop device-specific anomaly detection models utilizing variants of Recurrent Neural Network (RNN) autoencoders, such as LSTM autoencoders, GRU autoencoders, and bi-directional LSTM (BiLSTM) autoencoders. By employing RNN-based autoencoders for condition monitoring, it is possible to successfully identify abnormal behaviors in individual devices using a summarized signal. Moreover, the performance of different power disaggregation and anomaly detection models were compared separately, and the best-performing model for each device was discussed.