AA-LSTM: An Adversarial Autoencoder Joint Model for Prediction of Equipment Remaining Useful Life

摘要

Remaining Useful Life (RUL) prediction of equipment can estimate the time when equipment reaches the safe operating limit, which is essential for strategy formulation to reduce the possibility of loss due to unexpected shutdowns. This paper proposes a novel RUL prediction model named AA-LSTM. We use a Bi-LSTM-based autoencoder to extract degradation information contained in the time series data. Meanwhile, a generative adversarial network is used to assist the autoencoder in extracting abstract representation, and then a predictor estimates the RUL based on the abstract representation learned by the autoencoder. AA-LSTM is an end-to-end model, which jointly optimizes autoencoder, generative adversarial network, and predictor. This training mechanism improves the model's feature extraction and prediction capabilities for time series. We validate AA-LSTM on turbine engine datasets, and its performance outperforms state-of-the-art methods, especially on datasets with complex working conditions.