Document Type

Dissertation

Rights

This item is available under a Creative Commons License for non-commercial use only

Disciplines

1.2 COMPUTER AND INFORMATION SCIENCE, Computer Sciences, Information Science

Publication Details

Dissertation submitted in partial fulfilment for a M.Sc. in Computing (Data Analytics) School of Computing Dublin Institute of Technology

Abstract

Older IoT “smart sensors” create system alerts from threshold rules on reading values. These simple thresholds are not very flexible to changes in the network. Due to the large number of false positives generated, these alerts are often ignored by network operators. Current state-of-the-art analytical models typically create alerts using raw sensor readings as the primary input. However, as greater numbers of sensors are being deployed, the growth in the number of readings that must be processed becomes problematic. The number of analytic models deployed to each of these systems is also increasing as analysis is broadened. This study aims to investigate if alerts created using threshold rules can be used to predict network faults. By using threshold-based alerts instead of raw continuous readings, the amount of data that the analytic models need to process is greatly reduced. The study was done using alert data from a European city’s District Heating network. The alerts were generated by “smart sensors” that used threshold rules. Analytic models were tested to find the most accurate prediction of a network fault. Work order (maintenance) records were used as the target variable indicating a fault had occurred at the same time and location as the alert was active. The target variable was highly imbalanced (96:4) with a minority class being when a Work Order was required. The decision tree model developed used misclassification costs to achieve a reasonable accuracy with a trade-off between precision (.63) and recall (.56). The sparse nature of the alert data may be to blame for this result. The results show promise that this method could work well on datasets with better sensor coverage.

DOI

10.21427/D7CZ21

Share

COinS