Engine Failure Detection of Raw Mill Machine via Discrete Variational Auto-encoder

We present a deep learning model to detect failure engine state by observing the discrete latent sensor behaviors. Further, we investigate the behaviors from the reconstruction loss of the model until we find its value starting to jump out (anomaly stage). As a result, this work aims to forecast the failure time of the engine as early as possible. To notice the anomaly, we formulate a piecewise function based on
$\alpha$
-quantile of the loss value inside the proposed model. Unlike the existing studies focusing on the continuous latent, this work draws the discrete latent from discrete variational auto-encoder (DVAE) to predict the failure state better. For evaluation purposes, we evaluated the proposed model on a real dataset from the raw mill machine of a cement factory in Indonesia. From the experiments, we are satisfied to see the proposed model performances detecting the failure state of the raw mill machine as early as possible compared to the state-of-the-art model.