Conditional Variational Transformer for Bearing Remaining Useful Life Prediction

Transformer, built on the self-attention mechanism, has been demonstrated to be effective in numerous applications. However, in the context of prognostics and health management, the self-attention mechanism in the Transformer is not effective in selecting the most important features that are highly correlated with the remaining useful life (RUL) of a component. To address this issue, we developed a novel conditional variational transformer architecture consisting of four networks: two generative networks and two predictive networks. The first generative network uses the transformer encoder–decoder as well as both condition monitoring data and RUL as input to extract the most important features in one feature space from condition monitoring data. The second generative network uses the transformer encoder and condition monitoring data to extract features in another feature space. The two predictive networks use the extracted features in two different feature spaces to make predictions. A KL-divergence is used to minimize the distance between the two feature spaces learned by the first and second generative networks so that the feature space extracted from the second generative network can approximate the feature space extracted from the first generative network. We demonstrated that the proposed method is effective in predicting the RUL of bearings using two datasets.