A hybrid extreme learning machine and deep belief network framework for sludge bulking monitoring in a dynamic wastewater treatment process. (April 2022)

• Record Type: Journal Article
• Authors: Safder, Usman; Loy-Benitez, Jorge; Nguyen, Hai-Tra; Yoo, ChangKyoo
• Is Part Of: Journal of water process engineering. Volume 46(2022)
• Publication Date: 2022-04
• Subjects: Deep belief network -- Extreme learning machine -- Sludge bulking -- Sludge volume index -- Stacked-autoencoder; Water-supply engineering -- Periodicals; Saline water conversion -- Periodicals; Seawater -- Distillation -- Periodicals; Sanitary engineering -- Periodicals; Sewage -- Purification -- Periodicals; 627
• DOI: 10.1016/j.jwpe.2022.102580
• Languages: English
• ISSNs: 2214-7144

Abstract

Abstract: In biological wastewater treatment plants (WWTPs), sludge thickening is a common problem with major economic and environmental effects. Monitoring the sludge volume index (SVI) is a critical difficulty in sludge bulking quantification. This study develops a hybrid deep-learning-based soft sensor that uses a sparse constraint stacked autoencoder (SAE) model to estimate SVI and classify bulking states in WWTP effluent to improve effluent quality. The weights of the SAE in a training plan were adjusted using an extreme learning machine (ELM) technique. A deep-belief network technique was applied to the extracted feature space of the ELM to classify the current state of a secondary clarifier. The suggested methodology outperforms existing state-of-the-art methods in the SVI modeling task by 38% to 78%. On the other hand, the bulking state was properly classified with 90.7% accuracy and 91% precision. Graphical abstract: Unlabelled Image Highlights: An ELM-SAE model predicted the SVI of water quality with 98.39% accuracy. The feature space of the ELM was used to classify bulking sludge. A DBN was proposed to classify the current states of the secondary clarifier. The ELM-SAE and DBN efficiently matched non-linear dynamic systems. The proposed model proved high prediction accuracy compared with reported models.