Bayesian variational recurrent neural networks for prognostics and health management of complex systems

Abstract

In recent couple of years many automated data analytics models are implemented, they provide solutions to problems from detection and identification of faces to the translation between different languages. This tasks are humanly manageable but with a low processing rate than the complex models. However, the fact that are humanly solvable allows the user to agree or discard the provided solution. As an example, the translation task could easily output misleading solutions if the context is not correctly provided and then the user could improve the input to the model or just discard the translation. Unfortunately, when this models are applied to large databases is not possible to apply this double validation and the user might believe blindly in the model output. The above could lead into a biased decision making, threatening not only the productivity but also the security of the workers. Because of this, is neccessary to create models that quantify the uncertainty in the output. At the following thesis work the possibility to use distributions over single point matrices as weights is fused with recurrent neural networks (RNNs) a type of neural network which are specialized dealing with sequential data. The model proposed could be trained as discriminative probabilistic models thanks to the Bayes theorem and the variational inference. The proposed model is called Bayesian Variational Recurrent Neural Networks is validated with the benchmark dataset C-MAPSS which is for remaining useful life (RUL) prognosis. Also, the model is compared with the same architecture but a frequentist approach (single points matrices as weights), with different models from the state of the art and finally, with MC Dropout, another method to quantify the uncertainty in neuronal networks. The proposed model outperforms every comparison and furthermore, it is tested with two classification tasks in bearings from University of Ottawa and Politecnico di Torino, and two health indicator regression tasks, one from a commercial wind turbine from Green Power Monitor and the last in fatigue crack testing from the University of Maryland showing low error and good performance in all tasks. The above proves that the model could be used not only in regression tasks but also in classification. Finally, it is important to notice that even if the validations are in a mechanical engineering context, the layers are not limited to them, allowing to be used in another context with sequential data.