A Novel Multivariate Time-Series Anomaly Detection Approach Using an Unsupervised Deep Neural Network

Abstract

With the development of hardware technology, we can collect increasingly reliable time series data, in which time series anomaly detection is an important task to find problems in time and avoid risks. It is not easy to establish a multivariate time series anomaly detection system, because the collected data not only have different attributes, scales, and characteristic information but also have horizontal and vertical connections among these data collected by various sensors. In addition, there is no clear boundary regarding whether the data are abnormal, and there is currently no unified definition of multidimensional timeseries anomalies. Recently, deep learning methods have shown outstanding advantages in the processing of multidimensional time series. In this paper, we propose a definition of point anomalies in multivariate time series and an unsupervised deep learning method, the multilayer convolutional recurrent autoencoded anomaly detector (MCRAAD), which is used to detect anomalies in multivariate time series. We calculate the feature matrix sequence through the data in the sliding window, extract the characteristics of the feature matrix sequence by a multilayer convolutional encoder, obtain the time relations in the feature matrix by using several ConvLSTM units, and finally reconstruct the feature matrix sequence with the convolutional decoder to predict the self-feature matrix. In addition, we propose a threshold setting method to assist with the determination of anomalies. Finally, we test our model on synthetic datasets and a real dataset of house monitoring. The results of experiments show that our method is superior to these basic models in detecting ability and robustness. This model also provides an effective method for multivariate time-series anomaly detection in real life.

Online Multivariate Anomaly Detection and Localization for High-Dimensional Settings

“…The problem of anomaly detection has been an important subject of study in several research communities, such as statistics, signal processing, machine learning, information theory, and data mining, either specifically for an application domain or as a generic method. To name a few, an SVM classification approach for anomaly detection was proposed in [ 10 ]; Bayesian methods were developed for social networks [ 11 ], partially observed traffic networks [ 12 ], and streaming environmental data [ 13 ]; deep neural network models were proposed for detecting anomalies multivariate time series [ 14 , 15 , 16 , 17 , 18 ]; several information theoretic measures were proposed in [ 19 ] for the intrusion detection problem; and two new information metrics for DDoS attack detection was introduced in [ 3 ]. Due to the challenging nature of the problem and considering the challenges posed by today’s technological advances such as big data problems, there is still a need for studying the anomaly detection problem.…”

TCAD: Unsupervised Anomaly Detection Based on Global Local Representation Differences

“…However, Autoencoder cannot model temporal dependencies and has difficulty capturing anomalous sample contextual information. The LSTM Encoder-Decoder model proposes an Autoencoder model with LSTM as the backbone, which captures temporal context information by exploiting the modeling ability of LSTM for time.LSTM, although effective for temporal modeling, has the disadvantages of not being able to exploit spatial information and not being able to capture long time dependence.MCRAAD [31] proposes to convert The input sequence is transformed into a two-dimensional image, and the sequence is reconstructed with a three-layer two-dimensional convolutional encoder (CAE), using Convlstm for spatio-temporal information extraction in the middle of each layer, which effectively solves the drawback of LSTM's missing spatial information, but it lacks global temporal dependence. VAE [32] is a special Seq2Seq structure whose encoder learns a priori from the data X to learn a latent variable Z with multivariate Gaussian distribution N (0, I) a priori, and decoding then Z is sampled and reconstructed as X .…”

Recognition of wave characteristics in dredging construction area based on visual information

“…As for the prediction of ocean waves [6] , it is mainly based on large-scale wave models. This approach has several drawbacks: first, large numerical computational models, whose operation requires forecasters with certain model knowledge to operate [7] , and the existing research on ocean waves is mostly limited to the academic level [8,9] . Especially for near-shore waves, the influence of seafloor topography on waves is more complex [10] , and there are certain limitations in the near-shore wave height and period forecasts of numerical forecasting models [11] ; second, it is difficult to obtain detailed information on meteorology and geography needed to establish numerical wave prediction models for specific engineering range sea areas [12] , and it is difficult to establish high-precision numerical wave prediction models [13] .…”