Spatiotemporal Sequence Anomaly Detection Based On Spatiotemporal Embedding And Self-supervision

Mobile devices,the Internet of Things,and various sensor technologies are rapidly developing,resulting in massive amounts of spatiotemporal data being generated every day in the real world.This makes spatiotemporal data mining techniques more and more critical in various data analysis tasks.Anomaly detection is an essential part of this.Among them,an effective unsupervised spatiotemporal sequence anomaly detection system is very important for various decision-making tasks,including extreme weather warning,default prediction,fraud prevention,etc.Due to the complexity of the spatiotemporal sequence data,traditional anomaly detection algorithms are ineffective in detecting anomaly spatiotemporal sequences.Further,it is difficult to formulate accurate labels in the production environment,and the data collection process may generate substantial noise.Consequently,unsupervised techniques tend to be used in spatiotemporal sequence anomaly detection tasks rather than various data-driven supervised or semi-supervised approaches.At present,the vast majority of unsupervised spatiotemporal sequence anomaly detection methods use a hybrid architecture,that is,use deep neural networks to model spatiotemporal sequences and then design rules or other models to detect anomalies in the modeling results.The disadvantage of this type of method is that the anomaly detection target cannot be used for spatiotemporal sequence modeling,which makes the deep neural network can’t perform spatiotemporal sequence representation learning in the direction of anomaly detection.The performance of such methods is often sub-optimal and cannot perform end-to-end model training and anomaly inference.This paper studies the end-to-end unsupervised spatiotemporal sequence anomaly detection technology.The main work and innovations of the paper are as follows:1)The general unsupervised deep anomaly detection theory is studied,and a general unsupervised deep anomaly detection model named FDAE(Feature Decomposition AutoEncoder)is proposed.FDAE applied RPCA(Robust Principle Component Analysis)and RSR(Robust Subspace Recovery)technology at the feature level,making it robust to anomalous data of unknown categories and proportions.The model is compared with multiple state-of-the-art models on six public datasets,and the experimental results verify the effectiveness of FDAE;2)Combining the transformer architecture with FDAE,an unsupervised spatiotemporal sequence anomaly detection model named STADTrans(Spatial-Temporal Anomaly Detection Transformer,spatiotemporal anomaly detection transformer)is proposed,and a discrete expression of the spatiotemporal sequence is designed for it.STADTrans inherits the sequence modeling capability of the transformer and the anomaly detection capability of FDAE.With the aid of the spatiotemporal sequence discrete expression method designed in this paper,STADTrans can mine the normal patterns in the spatiotemporal sequence data;3)A self-supervised task—the accompanying reconstruction task is designed for STADTrans.Combined with the discrete expression of spatiotemporal sequences designed in this paper,this task transforms the spatiotemporal sequence anomaly detection problem into a multi-classification problem.STADTrans uses the classification error to calculate the anomaly score of spatiotemporal sequence.The accompanying reconstruction task unifies the spatiotemporal sequence modeling task and the anomaly detection target,enabling STADTrans to perform end-to-end unsupervised spatiotemporal sequence anomaly detection.This paper verifies the superior performance of the STADTrans model in the task of unsupervised spatiotemporal sequence anomaly detection on a large public AIS(Automatic Identification System)dataset.Then we further conduct performance tests on STADTrans in an actual production environment,and the results verify the practical value of STADTrans.