Comparison and extension of autoencoder models for uni- and multivariate signal compression in IIoT
Julia Rosenberger, Alexander Kubel, Fabian Rothfuss · 2022
A convolutional variational autoencoder (AE) CBN-$VAE^{1}$, a recurrent LSTM-$AE^{2}$, and a discrete wavelet transform (DWT) are compared w.r.t compression performance and resource consumption. The existing models for streaming data are slightly adapted to handle both univariate (UTS) and multivariate time-series (MTS). The experiments on two publicly available data sets3confirm that the machine learning models compress time-series in a more generalized and robust manner than DWT. This is shown by a higher quality score$QS= \frac{\text{compression ratio}(CR)}{\text{reconstruction error}(RE)}$for both AE models compared to the DWT. The following aspects are observed w.r.t. the AE models: