is a data mining task that aims to identify data that significantly differ from most other data. Many applications of anomaly detection in different contexts have been developed, including financial card fraud, industrial intrusion, and medical seizure detection. A large number of time series are recorded from a wide range of fields, including medicine, industry, and the natural sciences最新中文telegram的下载的网址在哪里. The analysis of time series with the aim of extracting useful information has become a major field of study. The extracted information can be used to solve a variety of problems. Anomaly detection on time series, as one of a problem, aims to discover the abnormal event from the time series recorded from variety areas. For example, Dwivedi et al. The paper [1] used an ensemble feature selection from a computer network and the grasshopper optimization approach to improve the detection of network intrusions through support vector machines. Siniosoglou et al. [2] used an autoencoder-based generative adversarial network to identify anomalies in a smart electrical grid system.
The tradition methods for anomaly detection are based on the statistical indices and density of the dataset. However, there some limitations for the tradition methods. For example, the performance of traditional methods in anomaly detection on medical images and sequential datasets is terrible because they cannot capture complex structures in the data最新的官网telegram下载网站. In addition, it is impossible for tradition methods to extent to large-scale data to find anomalies. Also, the boundary between the abnormal and normal behavior is often imprecisely defined in time series data and is constantly evolving. This lack of well-defined representative normal boundaries poses a challenge to traditional method. Therefore, the deep learning-based methods are proposed. Deep anomaly detection technology learns hierarchical discrimination features from the time series data. This automatic feature learning capability eliminates the need for domain experts to manually develop features. So that is advocated to solve the anomaly detection problem on time series.
There are lots of survey papers have introduced the deep learning methods for anomaly detection on time series. Such as the paper [3] reviewed the methods and applications related 官网的最新版telegram下载的地方在哪呢 detecting abnormal time series, and summarized the challenges for the methods. However, this paper only considers the one dimensional time series, and there are lots of multivariate time series are recorded from the real-world. In addition, the existing survey papers related to the time series anomaly detection did not introduce the commonly used datasets and the research challenge. Therefore, to fill these gaps, we reviewed the papers related to the methods, applications, research challenge, and datasets for anomaly detection on multivariate time series.
Because the input data and types of anomalies involved in the analysis of time series can vary considerably, distinct anomaly detection technologies are required. As shown in Fig. 1, there are three main types of anomalies, which are abnormal time point, time interval, and time series. The mainly applications based on the abnormal time point detection are fraud detection, attack detection, and noise detection. The abnormal time interval detection is commonly used for seizure detection, climate monitoring, and pollution monitoring. The faulty detection, state diagnosis, and event detection are based on the abnormal time series detection.
An abnormal time point is a certain time point at which the value of the monitored variable is significantly different from those at any other time points. This is formulated as or , where is the th time point, is the abnormal score for time , and is the threshold for detecting the anomaly. An abnormal time point in a time series generally corresponds to an extreme value. In some cases, the anomalies correspond to noise in the datatelegram精简版下载网址. Because in such cases the anomalies affect the data analysis results, they must be detected and removed from the given time series. In other cases, this type of anomalies may represent events of interest. For example, Fig. 2 shows three recorded electroencephalograms (EEGs) in which three extreme values are displayed. These abnormal points indicate either noise or a relevant event at that time point. Therefore, detecting this type of anomalies is useful for time series analyses. Ranjan et al. [4] improved the traditional prediction method based on time window sliding to detect anomalies in a given time series and increase forecasting accuracy for the planning and operation of a power system. Similarly, Wang et al. [5] applied a convolutional neural network (CNN) to detect anomalies in a time series collected from a wind power system to improve short-term prediction.
. Types of anomalies and applications.
. Example of an abnormal time point.
Abnormal time interval. An abnormal time interval is a certain time interval , during which the variable behaves significantly differently than during the rest of the time series. This is formulated as or , where is the th time interval, is the abnormal score for time , and is the threshold for detecting the anomaly, typically indicating the occurrence of an unusual event. Analysts may focus on these anomalies to explore the meaning of the detected eventstelegram最新官网下载的网站在哪里. Fig. 3 presents an example of an abnormal time interval in three EEG signals, where the red areas correspond to an epileptic episode suffered by a patient. Li and Jung [6] proposed an anomalous time interval detection method to identify seizures from multivariate EEG signals. Because brain discharge is quite different with and without a seizure, the authors considered the seizure interval as an abnormal time interval and proposed a graph-based detection approach. In this case, detecting this type of anomalies can help experts analyze and understand the patterns of abnormal events. Chen et al. [7] presented an approach for detecting network attacks using Internet of Things (IoT) time series. This study helped experts recognize and analyze the patterns of network attacks.
. telegram最新的官网的下载的网站是多少 Example of an abnormal time interval.
Abnormal time series. An abnormal time series is a certain time series , which is significantly different from other time series. This is formulated as or , where is the th time series, is the abnormal score for time , and is the threshold for detecting the anomaly. Fig. 4 shows three time series where two of the signals exhibit high correlation, in contrast with the one in red, which is thus detected as an abnormal time series [8]. Such as Li and Jung [9] discover the relationship between the stock and financial indies. Then, a dynamic graph is used for modeling these relationships. Finally, a abnormal stock is detected by proposing a graph embedding model.
. Example of an abnormal time series.
The remainder of this paper is organized as follows. In Sections 2, 3, and 4 technologies for abnormal point, abnormal time interval and abnormal time series detection are discussed, respectively. In Section 5, the approaches for selecting the threshold are discussed. In Section 6, applications based on anomaly detection in multivariate time series are reviewed. In Section 7, available open-access time series datasets for anomaly detection are listed最新官方telegram下载网址在哪呢. In Section 8, limitations and research challenges in anomaly detection in multivariate time series are presented. Finally, in Section 9, conclusions are discussed.