Anomaly Detection Based on Reconstruction Error in High-Frequency Hydroponic Sensor Data

In vertical farming and hydroponics, rapid changes in temperature, humidity, light, pH, and electrical conductivity raise the question of how to separate production-relevant deviations from normal operation. While previous studies have enhanced monitoring with IoT and machine learning, many approaches still rely on static thresholds and provide limited support for multivariate, high-frequency, and interpretable field monitoring. This study proposes an end-to-end framework that combines ESP32-based IoT monitoring with unsupervised anomaly detection for hydroponic vertical farming. The analyzed dataset contains seven consecutive days of synchronized 5-second measurements, corresponding to 120,960 raw records and 120,930 windowed sequences after 30-step sliding-window construction. The LSTM-Autoencoder learns normal multivariate temporal patterns and assigns each window a unified reconstruction-error score using mean squared error (MSE). Anomalies are flagged with a validation-derived statistical threshold (μ +2σ), resulting in a threshold of 0.0862 MSE and 3,454 detected anomalous windows (2.86%). To address interpretability, feature-wise reconstruction errors, event-level anomaly grouping, correlation comparison, and PCA of the LSTM latent representation are reported. The detected anomalies are short-lived and episodic rather than a sustained regime shift; 556 event-level groups were obtained, with a median duration of 25 s. Sensor contributions are broadly balanced, with humidity (20.22%), pH (20.18%), light (20.11%), EC (19.95%), and temperature (19.54%) showing no single dominant source. Additional threshold-sensitivity and controlled anomaly-injection baseline analyses clarify the robustness and limitations of the proposed pipeline. Overall, the framework is positioned as an explainable screening layer for anomaly prioritization rather than a fully validated root-cause decision system.