Unsupervised Anomaly Detection in Hospital Wastewater Effluent Using Convolutional Autoencoder

Abstract

Background: Hospital wastewater treatment plants (WWTPs) play a crucial role in maintaining environmental sustainability. However, conventional monitoring has difficulty identifying minor differences in effluent quality, leading to non-compliance. While machine learning is increasingly applied in water quality analysis, the specific application of deep representation learning in hospital effluent analysis, focusing on identifying anomalies within stable and low variation factors, is not much explored. Objective: This study aims to evaluate the effectiveness of a proposed Convolutional Autoencoder (Conv-AE) for anomaly detection in the effluent of hospital WWTP. To ensure the efficacy of the algorithm, it is compared with two popular statistical algorithms: Isolation Forest (IF) and One-Class Support Vector Machine (OCSVM). Methods: Internet of Things (IoT) sensor data covering pH, temperature, Total Dissolved Solids (TDS), and ammonia gas parameters were collected from the effluent tank of a hospital WWTP. The Conv-AE model was designed to learn the latent nonlinear representations of normal effluent patterns. The model’s performance was evaluated using precision, recall, F1-score, accuracy, and inference time metrics. Results: The proposed Conv-AE model performed best in terms of detection, having the best values ​​for all three metrics, with a recall of 0.980, an F1 score of 0.960, and an accuracy of 0.980. This indicates a robust ability to identify subtle deviations that statistical baselines often miss. In terms of operational feasibility, while the Isolation Forest baseline exhibited the fastest inference time of 0.000014 seconds, the Conv-AE remained highly efficient for real-time applications with a inference time of 0.000348 seconds. Conclusion: In conclusion, the Conv-AE algorithm offers an optimal trade-off between high detection sensitivity and operational feasibility. By prioritizing the minimization of false negatives, this deep learning approach provides a more reliable solution for safety-critical hospital effluent monitoring compared to traditional statistical partitioning methods. Keywords: Anomaly Detection, Hospital Wastewater Treatment Plant (WWTP) Effluent, Convolutional Autoencoder, Deep Learning