Gaussian process latent variable model and Bayesian inference for non-parametric failure modeling applied to ship engine

This is the authors' abstract. We don't add key points for this paper.

• Presenting a Gaussian Process-based latent space model combined with Bayesian inference for non-parametric, time-dependent failure analysis. • Demonstrating the methodology's applicability for predictive maintenance and safety analytics in the context of an unattended machinery plant, paving the way for Maritime Autonomous Surface Ships (MASS). • Reaching remarkable accuracy with 96.5% of the predictions aligning with observed data, as validated by Pareto k-diagnostic values, highlighting the model's reliability and well-calibrated predictive distribution. Unnecessary early maintenance is especially critical for high-value or essential components whose unexpected failures could disrupt the entire operational process of the system. The uncertainties inherent in facility deterioration necessitate a robust framework that accurately assesses system health and guides optimal maintenance scheduling. To this end, this paper proposes a probabilistic machine learning framework based on a Gaussian Process Latent Variable Model (GPLVM) combined with Bayesian Inference (BI) to dynamically assess the health state of system and predict failure risk. The model integrates uncertainty quantification through BI, providing a non-parametric hazard rate estimate at each time step, which enables a precise and adaptive maintenance planning strategy. To verify the proposed model, a critical component of an engine – spark ignition, is considered as the case study. Herein, ignition voltage is monitored as the primary indicator of spark health, with degradation thresholds and safety thresholds explicitly modeled to capture degradation trends accurately. The results indicate that 96.5% of the observations fell within precise predictive range (according to Pareto Diagnostics values), underscoring the model’s promise for maintenance planning. This approach has the potential not only to improve predictive accuracy and decision confidence but can also provide a flexible, non-parametric solution adaptable to various high-stakes maintenance applications.