Latent approaches for unsupervised learning-based methodologies in gearbox fault diagnosis applications

Abstract

The monitoring of critical industrial assets is of fundamental importance to the longevity and sustainability of many industrial sectors. Vibration-based fault diagnosis methodologies seek to enhance the monitoring process by providing rich diagnostic information related to the instantaneous asset health state to support maintenance decisions. While conventional signal analysis techniques applied to vibration data have been demonstrated to provide information related to the asset health state, their utility is often impeded by other vibration sources and time-varying operating conditions. Hence, powerful data-driven techniques are used to extract asset health state information to support the decision-making process.

Learning-based methodologies offer the ability to perform effective fault diagnostics by extracting informative features for the fault diagnosis problem. Unsupervised learning latent-based approaches, e.g., latent variable models (LVMs), offer an effective strategy to bypass dataset-dependent feature engineering practices and the requirement for labelled fault conditions inherent to many learning-based methodologies. Temporal preservation is an important approach which advocates for the preservation of time structure to produce an informative latent manifold, and new possibilities and opportunities for condition inference are potentially available from a latent variable perspective. However, this is only the first step to unlocking the potential of LVMs for asset condition monitoring. Improving the application of temporal-preserving LVMs is crucial to advance unsupervised learning-based methodologies for effective utilisation and adoption in industrial applications. The approaches considered in this thesis address current issues in gearbox condition monitoring, and areas for improvement are identified for temporal-preserving LVMs. Three research directions are established to improve the application of temporal-preserving LVMs for gearbox fault diagnostics. The first direction focuses on leveraging the latent manifold to drive effective condition inference, the second direction explores different LVM training approaches, and the third direction investigates latent health indicators for LVMs developed using different training approaches.

First, a latent indicator framework for temporal-preserving LVMs is proposed, which identifies unique classes to capture several manifold perspectives for latent health indicator derivation. The benefits of an ensemble-based latent condition inference approach are identified, and the latent manifold is demonstrated to be a fruitful source for fault diagnostic information. This enhances LVM-based methodologies for the fault diagnosis task. As the models are fault condition agnostic, it is demonstrated that a diverse set of latent health indicators is necessary to obtain a robust and insightful condition inference approach.

Second, the parameter estimation process for temporal-preserving LVMs is studied to explore methods used to construct the latent manifold. For data-driven models in condition monitoring, two strategies exist to estimate the model parameters: offline and online training. Albeit equally applicable for LVMs, conventional research applications disregard online strategies for LVMs. Online training suits scenarios which lack historical data from a healthy asset. Thus, understanding the implications of using online training for temporal-preserving LVMs is considered. Moreover, two main methods are identified for online training: reconstruction-focused and interpretation-focused methods. Both online and offline models are demonstrated to be effective using basic methods from the proposed latent indicator framework, which reveals their utility in condition monitoring and confirms their ability to perform effective fault diagnostics.

Finally, the proposed latent indicator framework pre-supposes an offline training approach and was only assessed using offline models. Therefore, the latent indicator framework bounds are studied using online models. Implicit assumptions in the proposed latent indicator framework are identified to develop a new category of latent health indicators, and a set of online latent indicators is proposed for online methods. The applicability of the two indicator categories is studied. Diversity is highlighted to be a prominent factor in the indicator categories, and the study advocates for the use of a plethora of diverse indicators for both offline and online methods.

The research carried out in this thesis indicates that the considered temporal-preserving latent variable approaches are advantageous to the fault diagnosis task. The identified research directions ensure that the latent manifold of offline and online temporal-preserving LVMs can be effectively applied to drive meaningful condition monitoring. By combining offline and online models with a diverse, categorised latent indicator framework comprising of heterogeneous collection latent health indicators, a holistic fault diagnosis strategy for temporal-preserving LVMs can be developed. Temporal-preserving LVMs are established as an effective tool for unsupervised learning-based condition monitoring using gearbox vibration data.