Machine learning - Gaussian-Process

Gaussian Process (GP) models are well suited to small or noisy datasets and provide an uncertainty estimate alongside each prediction. We can build GP classifiers or regressors for tasks such as fault status assessment, remaining-useful-life estimates, or anomaly scoring. We focus on models that engineers can act on: we show predicted values and confidence, and we explain how thresholds map to false alarms and misses. When data are scarce, we can set up “active learning” so testing effort is focused where it reduces uncertainty the most.

In multi-partner projects, this helps analytics work stay robust and transparent. GPs integrate well with physics-based features, so results can be cross-checked against models rather than treated as a black box. We document assumptions and provide compact notebooks and scripts partners can run.

Example applications
• Classifying hatch state from multiple weak sensor cues.
• Estimating a condition index with confidence intervals for maintenance decisions.
• Ranking test points or components by “information gain” to plan the next experiment.

Deliverables
• Clean training and evaluation notebooks, model summaries, calibrated thresholds.
• Plots and short reports showing performance and where uncertainty remains.