Abstract:
Latent variable models are important for condition monitoring as they learn, without any supervision, the healthy state of a physical asset as part of its latent manifold. This negates the need for labelled fault data and the application of supervised learning techniques. Latent variable models offer information from which health indicators can be derived for condition monitoring. Namely, information from the latent space and the data space can be used for condition inference. These health indicators are used to explain changes in a physical asset’s condition. Conventional black-box approaches only offer information from the data space in the form of reconstruction errors. In contrast, latent variable models offer a latent space and reconstruction space for inference. However, the current application of latent variable models either disregards latent space information or fails to realise its full potential. The full potential can be realised by preserving the time information in the data. Therefore, we propose a model evaluation procedure that specifically preserves time in the latent health indicators. The procedure is generic and can be applied to any latent variable model as demonstrated for Principal Component Analysis (PCA), Variational Auto-Encoders (VAEs) and Generative Adversarial Networks (GANs) in this study. In general, as time information can be discarded or preserved for derived latent health indicators, this study advocates that health indicators that preserve time are more useful for condition monitoring than health indicators that discard time. In addition, it enables the interpretation of the learnt latent manifold dynamics and allows for alternative latent indicators to be developed and deployed for fault detection. The proposed temporal preservation model evaluation procedure is applied to three classes of latent variable models using two datasets. Three model-independent latent health indicators that preserve time are proposed and shown to be informative on all three classes of latent variable models for both datasets. The temporal preserving latent analysis procedure is demonstrated to be essential to derive more informative latent metrics from latent variable models.
