Lab-ver: an LSTM attention based on variational autoencoder representation learning of remaining useful life estimation

Abstract A representation learning model based on a variational autoencoder(LAB-VER) was introduced for the remaining useful life (RUL) prediction task. Deep learning methods are effective in learning the features within the data and leveraging these features to conduct tests and evaluations on the equipment. However, existing methods often lack the ability to represent data features effectively. To address this limitation, a novel representation learning model based on a variational autoencoder was proposed. The model consists of an encoder and a regression: the encoder learns to decouple the input data into the feature space, while the regression maps the decoupled data features to obtain the prediction results. Within the encoder of the VAE, a parallel information fusion and enhancement approach has been developed, ingeniously integrating the distinct modules of LSTM, multi-head attention, and nonlinear transfer mapping. Within the architecture, a compact fusion sub-module is employed. This approach leverages the processing results of each layer, thereby effectively enhancing the decoupling capability of the encoder component. On the C-MAPSS simulated dataset, the model outperforms most state-of-the-art methods, achieving root mean square errors of 10.10, 10.36, 8.73, and 9.85 on the four sub-datasets, with RUL prediction scores of 231.10, 755.01, 202.62, and 800.79, respectively.