Online Aware Synapse Weighted Autoencoder for Recovering Random Missing Data in Wastewater Treatment Process

Abstract

Missing values in wastewater treatment process (WWTP) data hinder the monitoring and prediction of operational status. Therefore, various online imputation methods have been proposed to recover missing values from streaming data collected in WWTP in real time. However, existing methods tend to ignore previous learned knowledge. In this paper, an online aware synapse weighted autoencoder imputation method (OASI) is proposed to impute random missing values. First, an online stacked autoencoder (OSAE) framework is constructed to capture the nonlinear structure of the recently collected data. The framework decreases the computational and storage consumption of the model training. Second, an aware synapses weighted parameter regularization strategy is presented to guide the update of model parameters and alleviate the forgetting of historical information in an online continual setup. In this way, the learned features offer a more comprehensive representation of the overall information and help enhance imputation performance. Third, two real WWTP datasets with strong non-stationarity, high-noise level and high-dimensionality are used to validate the performance of the proposed OASI. Experimental results show that the proposed OASI achieves superior performances over the existing methods even in the presence of random missing values with different missing ratios, and only costs a short running time.