Abstract:
Condition-based monitoring for critical assets is reliant on the quality of the indicators used for condition inference. These indicators must be sensitive to the development of faults under constant and non-stationary operating conditions. Latent variable models in the time-preserving framework offer a powerful learning-based technique for the monitoring of critical assets as they only require healthy asset data, and provide indicators from the data space for asset condition inference. However, the latent manifold of latent variable models is often disregarded in favour of data space indicators, such as reconstruction errors, which are primarily focused on measuring the likelihood of the observed data. The latent space is often unintentionally unutilised. In this work, we highlight that the latent manifold is a powerful resource for condition inference and should be utilised for condition monitoring. We conceptually categorise and identify five classes of latent space health indicators that capture various manifold perspectives. These five classes introduce and allow for latent health indicator derivation and are useful for the condition inference task. Fifteen latent health indicators are considered in this work and are applied to the fault diagnostics task on two experimental datasets. An ensemble-based inference procedure is used, which produces a modular fault diagnosis framework. The indicators are shown to be informative for condition inference in both constant and variable operating conditions. Utilising the latent manifold in condition monitoring tasks is important to further develop the learning-based condition monitoring field.
