Citation

Mohd Zebaral Hoque, Jesmeen and Hossen, Md. Jakir and Abd. Aziz, Azlan (2022) Unsupervised Seasonal Novelty Detection for Yahoo Benchmark Data using Deep Learning. In: 2nd FET PG Engineering Colloquium Proceedings 2022, 1-15 December 2022, Multimedia University, Malaysia. (Unpublished)

Text 14_Jesmeen__PPM Extended Abstract_Second Half 2022.pdf Restricted to Registered users only Download (106kB)

Abstract

Background – In the current IoT and smart technology era, a tonne of data is accessible for analysis and decision-making. With the Smart Home system (SHS), an integrated Anomaly Detection (AD) model is also required to identify unexpected behavior and clean out unusual data. There hasn't been much research on novelty AD and handling unlabeled datasets. Purpose – The developed approach will be able to train the structured learning approaches. The system will be able to extract features from time-series data. In this study, a method for identifying abnormalities in univariate time-series data was developed and integrated into a function that any SHS programmer can also use. Additionally, the algorithm will tell a seasonal abnormality apart from a true abnormality. Design/methodology/approach – This procedure trains the learning model using the normal value of the time series and unlabeled generated data. Long short-term memory (LSTM) and deep convolutional neural network (CNN) were used to train the forecasting time series module. The trained AD model was obtained to detect abnormalities in real time. The trained model was loaded into the practical implementation. The procedure estimates the subsequent values of the selected window by analyzing a variety of windows of series data. Next an algorithmic process, it is determined whether closely the predicted values match the actual values. The data is divided between normal and abnormal data based on the threshold (a distance value) and abnormality fraction percent. Findings – The algorithm will be capable of working without the need for a labeled dataset by using the unsupervised technique. The model will first be learned using regular data, and thereafter predictions will be made by inputting anomaly datasets, enabling novelty AD. Finally, the method has been evaluated using yahoo benchmark datasets. Research limitations– The deep learning models cannot distinguish noise as abnormal data. Hence, to initially train the model it was important to clean the data. Originality/value – There have been studies on AD. Very few works done related to seasonal novelty AD. This research analyzed seasonal-based detection for unknown new abnormalities unknown in the dataset.

Item Type:	Conference or Workshop Item (Other)
Uncontrolled Keywords:	Anomaly detection, deep learning, feature extraction, time series, novelty
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	16 Feb 2023 06:40
Last Modified:	16 Feb 2023 06:40
URII:	http://shdl.mmu.edu.my/id/eprint/10834

Downloads

Edit (login required)