Deep learning-based production assists water quality warning system for reverse osmosis plants

Abstract

Classifying water quality irregularities in reverse osmosis (RO) production plants requires suitable systems to provide intelligent warnings to the operators or supervisors who are engaged in executing corrective procedures applicable to production. The suggested deep learning methods are of utmost importance to identify at once variations in water quality irregularities in plants through competent classification methods, thereby enabling a reduction of burden for operators. In this paper, two types of LSTM-CNN based classification techniques are suggested to classify water quality features temporally into grades based on corrective actions that classify irregularity conditions of water quality on the basis of corrections. Distinct control methods are used for experiments to find water quality irregularities from variables, namely, pH, TDS, ORP, and EC, which aim to assist the production line. This proposed method enables automatic diagnosis and warning systems about water quality in RO plants. For classification, LSTM-CNN was trained with data recorded from eight plants of west and north parts of Chennai region. This research is meant to demonstrate particularly the top-level classification job for quality alerts. The features obtained from 4,096 time series array data using LSTM-CNNs achieved sensitivity to 97% and accuracy to 98%.

Assessing the Performance of Deep Learning Algorithms for Short-Term Surface Water Quality Prediction

“…This is particularly true for those who take advantage of better performance from deep learning than its traditional counterparts such as machine learning and statistical models [1,4,5]. The research applied to hydrologic and water quality (time series) data ranged from flood and run-off forecasting through water quality and quantity modeling to modern chemical process, fisheries, and aquacultural engineering, just to name a few [6][7][8][9][10]. Despite its potential advantages, the performance of deep learning was found to be highly sensitive to the number, size, and type of layers, and to a less obvious extent, loss functions, optimization procedures, and so on [11,12].…”