Markov chain-based feature extraction for anomaly detection in time series and its industrial application
Dong Zang, Jinhai Liu, Huaizhen Wang · 2018
Different time series often represent different meanings, especially when time series are anomalous, carrying out the research of time series is of great significance. Unlike traditional anomaly detection methods, we propose a special feature extraction method based on transfer probability of Markov chain, so as to deep mining information of the raw time series. Firstly, time series is transformed into forms of Markov chain by a simple rule. Then, the transfer probability matrix of Markov chain can be calculated, we name it as Markov feature. Importantly, key parameters occurred during process of features extraction will be studied. And then, taking the extracted Markov feature and two simple statistical features(mean and variance) as the feature vector. Last, to verify effectiveness of the proposed method, we use four machine learning methods(i.e. k-NN, RF, DT and SVM) and oil pipeline pressure data in experiment section. Meanwhile, comparatives between the proposed features extraction with other three methods are done. Experimental results show that the proposed anomaly detection method is more accurate and effective.