Data-driven-modeling-based sensor placement for the detection and identification of failures in cyber physical water distribution networks.

Author

Utsav Parajuli, Southern Illinois University CarbondaleFollow

Date of Award

5-1-2023

Degree Name

Master of Science

Department

Civil Engineering

First Advisor

Shin, Sangmin

Abstract

The advance in Information and Communication Technology (ICT) has led to the transition of conventional water supply systems into water cyber-physical systems (CPSs). The water CPSs consist of the Supervisory Control and data acquisition (SCADA) system, making it vulnerable to various cyber-physical intrusions. In addition, water CPSs routinely experience anomalies from other conventional physical and operational failures. Although several data-driven methods have been created and validated for detecting a specific, single failure type, a comprehensive framework to distinguish and identify the failure type from other various types of possible failures is needed. In this context, this thesis addresses the following questions: 1) can the existing data-driven machine learning models contribute to the descriptive classification of failure types?; 2) is a sensor placement created for identifying a certain failure type also effective in identifying other failure types?; 3) does the sensor placement that is selected optimally for a particular failure type provide good performance for real world data under the data noise?. To find the answers and test the corresponding hypotheses, this study presents sensor placement with an integrated framework of unsupervised and supervised data-driven models for failure detection and identification. The integrated framework is applied to C-town WDN. As the first part, the integrated framework is designed as the combination of a deep learning autoencoder (AE) model and supervised multiclass classification models - Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Networks (ANN). The AE model is used for training the normal dataset with continuous SCADA features to detect unknown failures based on the threshold of the reconstruction error. This AE model showed acceptable performance to detect anomalies occurring from a certain failure event among various types. In identifying and differentiating the failure event among various types, SVM model demonstrated the best performance among three supervised failure identification models- SVM, RF, and ANN. As the second part, this study proposed a feature selection approach for the integrated failure detection/identification framework in the first part, which suggests the best sensor placement strategy for classification-based failure identification in a water CPS. The approach adopts a wrapper-based feature selection method that employs recursive feature elimination (RFE) with an RF classifier for feature selection. The performance of sensor placement options in failure detection and identification is evaluated using deep learning (AE) feature extraction followed by a RF model. The primary finding indicated that the best sensor option targeting a certain, single failure type shows a little bit lower performance in differentiating a failure event from various types. This provides useful insights to engineer water CPSs’ sensor placement strategies to detect, identify, and differentiate a failure event from various possible failure events. Also, the findings suggested that the performance of the integrated framework for failure identification has better performance under data noise when the embedded data-driven models are trained by realistic failure datasets including data noise.