From logs to maintenance decisions: explainable transformer-based anomaly detection and subsystem prioritization at Paranal Observatory

Abstract

Large astronomical observatories depend on complex software and hardware infrastructures whose behavior is continuously recorded through operational logs. These logs contain valuable information about abnormal events, degraded performance, and subsystem interactions, but their volume and heterogeneity make manual inspection difficult.

At ESO Paranal Observatory, this challenge is especially relevant because the Very Large Telescope and the Very Large Telescope Interferometer generate large-scale logs that must be interpreted to support timely maintenance decisions.

This paper takes a deliberately operational view of explainable log analytics: rather than emphasizing model derivations, it focuses on how anomalous log messages can be turned into concrete maintenance priorities for the people who run the telescopes. The pipeline classifies individual logs as normal or anomalous with a sentiment-aware Bidirectional Encoder Representations from Transformers (BERT) model, uses SHapley Additive exPlanations (SHAP) to extract the terms driving anomalous classifications, aggregates this evidence into a subsystem-instrument decision matrix, and ranks subsystem criticality with an extended Multi-Objective Optimization based on Ratio Analysis plus the full Multiplicative form (MULTIMOORA) method.

Applied to more than 5.7 million Very Large Telescope Interferometer logs from GRAVITY, PIONIER, and MATISSE at processing rates above 8,000 logs per second, the framework identifies ARAL, ISS, and RMNREC as the most exposed subsystems, distinguishes broad subsystem-wide degradation from localized instrument-subsystem interactions, and yields stable rankings across alternative weighting and ranking schemes. We translate these outputs into an actionable maintenance playbook that covers inspections, recalibration, spare provisioning, root-cause analysis, and daily triage, in which each priority remains traceable to the specific terms and log messages behind it, and we are explicit about the boundary between methodological stability and true operational validation against maintenance records.