Univariate anomaly detection in pressure data of a pilot aerobic membrane bioreactor unit using long short-term memory autoencoders.

This work presents a novel approach for univariate anomaly detection in membrane pressure of an aerobic membrane bioreactor (aMBR) using Long Short-Term Memory (LSTM) autoencoders. Membrane fouling, manifested by an increase in applied or transmembrane pressure, is one of the most critical challenges in the operation of aMBRs, as it directly affects effluent quality, energy consumption, and operating costs. Early and reliable fouling detection is therefore essential for maintaining process stability and reducing downtime. The study utilizes experimental data from an aMBR pilot plant treating Pulp and Paper (P&P) effluents, where applied pressure is a key indicator of membrane fouling and operational performance. Aiming to enhance real-time monitoring and increasing operational efficiency, the proposed methodology involves training an LSTM autoencoder to capture the normal temporal patterns of applied pressure and detect deviations that indicate anomalies such as fouling development, or other operational issues. The model exhibited high performance with R2 = 0.996 in reconstructing previously unseen normal data. Furthermore, the model's efficacy was tested using case studies simulating sensor malfunctions and irreversible membrane fouling, where the model successfully identified these anomalies in real time. In contrast to prior studies that primarily relied on synthetic or laboratory datasets, this work uses real pilot-scale operational data, thereby providing insights that are more representative of system behaviour. The performance of the proposed model suggests that LSTM autoencoders can be an efficient tool for predictive maintenance and anomaly detection in aMBR systems, paving the way for better management and optimization of wastewater treatment processes.