IoT-Based Unsupervised Anomaly Detection for Multivariate Time-Series Sensor Data in C. vulgaris Cultivation

Photobioreactor (PBR) microalgae cultivation involves complex multivariate physicochemical interactions in which subtle disturbances may affect multiple parameters simultaneously before visible culture degradation occurs. Conventional monitoring approaches based on manual sampling and univariate threshold inspection are often inadequate for capturing these coupled temporal dynamics in cultivation environments. However, the relative effectiveness of different unsupervised anomaly detection (AD) methods for modeling complex multivariate environmental data remains insufficiently understood, highlighting the need for a systematic comparative evaluation. This study aims to evaluate three unsupervised AD models, namely Isolation Forest (IForest), One-Class Support Vector Machine (OC-SVM), and Local Outlier Factor (LOF), for monitoring C. vulgaris PBR cultivation using multivariate real-time IoT sensor data. The observations comprising oxidation-reduction potential (ORP), electrical conductivity (EC), potential of hydrogen (pH), and water temperature were collected and analyzed as multivariate time-series data. Percentile-based thresholds (P90, P95, and P99), derived from the normal training data, were evaluated to examine the trade-off between anomaly sensitivity and false alarm reduction. Within the threshold-sensitivity analysis, IForest achieved a recall of 0.9750 and an F1-score of 0.9656 at P95, together with an ROC-AUC of 0.9957 and a PR-AUC of 0.9959. Temporal anomaly profiling further revealed that anomaly clusters were concentrated during the adaptation and stationary growth phases. These findings demonstrate the potential of unsupervised AD models as practical tools for data-driven monitoring and anomaly profiling in microalgae PBR cultivation systems.