Research On Outlier Detection In Multidimensional Time Series And Its Application

Outlier detection aims at detecting the outliers which generated from abnormal patterns.Multidimensional Times Series(MTS)is a collection of observations for multidimensional variables recorded in chronological order.The driving data is a kind of MTS and outlier detection for driving data is important to ensure life safety.Most existing detection models cannot detect outliers in driving data accurately.Therefore,the paper proposes two outlier detection models to detect outliers in driving data and flight data.QAR(Quick Access Recorder)data belongs to flight driving data.This paper focuses on the outlier detection in driving data and QAR data.Specific work includes:Firstly,the paper proposes an outlier detection model based on hybrid method,named LAIF(Long-short-term-memory based Autoencoder with Isolation Forest).The model combines deep learning and isolation forest to detect outliers in driving dataset.Sliding windows are used to complete the feature extraction from the original data and then an autoencoder based on Long Short Term Memory(LSTM)is used to reconstruct the feature data which makes the outlier easier to detect.After that the LAIF exploits isolation forest to finish detecting outliers.The experimental dataset is a public driving dataset—UAH-DriveSet.Comparing to baseline models,the experiment results show that the performance of LAIF in outlier detection has been significantly improved.Secondly,based on the LAIF model,LAE(LSTM based Auto Encoder)is proposed to deal with outlier detection in QAR data.The model contains two parts,feature extraction and outlier detection.The calculation in feature extraction is the same as LAIF.Sliding windows are used to extract data fluctuation features and then calculates data statistical information to complete feature extraction.Because of the excellent ability in learning hidden features,autoencoder is employed in training normal data.The LAE learns the features of normal data.The core of the LAE in detection is that features in outliers are different from normal data.When input outliers into the LAE,the output would deviate far from the input.The experiment results show that the LAE have better performance in outlier detection than other baseline methods.