Machine learning approaches for anomaly detection of water quality on a real-world data set

• Published in: Journal of information and telecommunication (Print) Vol. 3; no. 3; pp. 294 - 307
• Main Authors: Muharemi, Fitore, Logofătu, Doina, Leon, Florin
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis 03-07-2019; Taylor & Francis Ltd; Taylor & Francis Group
• Subjects: Algorithms; Anomalies; Artificial neural networks; Classification; Computer simulation; Discriminant analysis; Drinking water; event; F1 score; imputation; Learning theory; Machine learning; Neural networks; Performance evaluation; Recurrent neural networks; Regression analysis; Sensors; Support vector machines; Time series; Water quality; Water supply
• ISSN: 2475-1839; 2475-1847
• DOI: 10.1080/24751839.2019.1565653

Accurate detection of water quality changes is a crucial task of water companies. Water supply companies must provide safe drinking water. Nowadays in different areas, we find sensible sensors which monitor data during the time. Normally the data registered by the sensors contain a meaning, such as there can be any event. Sometimes the data are ill-understood and stating if there is an event which is difficult. This work represents the description of several approaches to identifying changes or anomalies occurring on water quality time series data. This work also discusses and proposes a solution to some challenges when dealing with time series data. The following models are applied to water quality data: logistic regression, linear discriminant analysis, support vector machines (SVM), artificial neural network (ANN), deep neural network (DNN), recurrent neural network (RNN) and long short-term memory (LSTM). The performance evaluation is conducted using F-score metric. A simulation study is conducted to check the performance of each algorithm using F-score. Solving imbalanced data is basically intentionally biasing the data to get interesting results instead of accurate results. The results show that all algorithms are vulnerable although SVM, ANN and logistic regressions tend to be a little less vulnerable, while DNN, RNN and LSTM are very vulnerable.