Abstract

A data-driven monitoring workflow is proposed to support operational interpretation of longitudinal photobioreactor datasets collected during microalgal cultivation in a lab-scale flat-panel system under continuous operation. Biomass (g/L) and nitrate (mg/L) were measured offline at regular campaign checkpoints, while key operating variables (e.g., optical signals, pH, temperature, gas flow, dilution rate, and harvesting/volume records) were logged to contextualize process evolution. The workflow targets monitoring and event surfacing rather than mechanistic attribution or long-horizon forecasting. It integrates (1) change-point detection to segment the time series into statistically distinct operating regimes, and (2) an ensemble anomaly detector (Isolation Forest + Local Outlier Factor) applied to a causal feature representation to rank observations by atypicality and highlight candidate excursions (e.g., short-lived perturbations, measurement irregularities, or uncommon operating states). To avoid over-reliance on a single hyperparameter choice, robustness is quantified in a reference-free sensitivity grid over segmentation penalties and anomaly contamination levels, yielding consensus regime boundaries and stability diagnostics for anomaly rankings without requiring event labels. Results show that nitrate dynamics are characterized by a small set of highly reproducible regime transitions, whereas biomass exhibits a multi-scale structure with both high-support boundaries and penalty-dependent finer partitioning. In both variables, anomaly-score rankings remain stable across plausible settings. The proposed workflow provides a transparent, uncertainty-aware monitoring layer that compresses raw trajectories into actionable regime markers and prioritized excursions, enabling targeted review and facilitating future integration with predictive or hybrid digital-twin modules.