Search for anomalies in linguistic models of time series

Abstract

To date, anomaly detection is one of the main reasons for performing data analysis. With the further development of the Internet of Things, the need for automated monitoring and decision-making systems capable of timely identifying disruptions or malfunctions of various devices and infrastructure and avoiding undesirable effects will only increase. That is why this article focuses on the development of effective anomaly detection algorithms. Practical results of the analysis of time series of stock prices of world famous campaigns are presented. When used in the analysis of time series, the term "anomaly" refers to the presence of values that are significantly distinguished from the general pattern, that is, values, in particular, values that are not due to the trend, cyclical nature of the series or seasonal bursts. The linguistic method of anomaly detection is to apply linguistic process modeling. The main purpose of linguistic modeling is to convert numerical series, experimental, multidimensional data to linguistic sequences and to derive a formal grammar of the language of the appropriate nature to solve the following range of problems: analysis and prediction of time series, recognition of patterns of various nature, authentication of the user, authentication state of the operator, diagnosis of diseases of the musculoskeletal system of operators of complex technical systems in the early stages of the disease ting. Experimental studies were performed using the described software. The experimental set of time series of securities index of the 30 largest US companies - the Dow Jones Industrial Average was used. During the development of the algorithm, 4 different metrics for evaluating linguistic models were considered. Given their strengths and weaknesses, root mean square was chosen as the most appropriate approach. 7 experiments were performed with the above software implementation of anomaly search algorithm based on linguistic time series modeling. The experiments confirmed the effectiveness of the method, as all anomalies were accurately detected for 5 of the 7 series studied.