| With the application of various sensors in recent years,a large number of spatio-temporal data is generated.It is of great practical significance to analyze data characteristics and detect anomalies based on sparse spatiotemporal data.Researchers have made some achievements in the field of anomaly detection of sparse spatio-temporal data,including distance-based anomaly detection methods,density-based anomaly detection methods and machine-learning-based anomaly detection methods.However,most of the existing anomaly detection methods usually reconstruct the data in order to overcome the problem of data sparsity.Data reconstruction will consume a lot of extra computing resources,and it is likely to produce redundant data,which is not conducive to anomaly detection.To solve the problem of anomaly detection on sparse spatio-temporal data,this thesis studies the anomaly detection methods of sparse data in spatial and temporal dimensions.Based on the sparsity of texture features of data images,combined with machine learning methods,the spatial anomaly detection is carried out.Aiming at the sparsity of time dimension,the deep learning methods are improved to detect spatio-temporal anomaly.The spatio-temporal data anomaly detection systems are designed and implemented.The main work of this thesis is as follows:1.Aiming at the sparse data images without time label,an anomaly detection method based on texture dictionaries is designed from the spatial dimension.In this method,Gabor filter is used to extract the texture features of the data images,and the texture dictionaries are constructed.On the basis of the complete texture dictionaries,the clustering algorithm is used to construct the corresponding index dictionaries to improve the detection efficiency.The detection accuracies of three experimental data sets based on this method are higher than 80%.And the false detection rates are low.2.For sparse spatio-temporal data sequence with time label and its generated images,an improved anomaly detection method based on generative countermeasure network is proposed.This method combines the time dimension and space dimension.An U-Net network is uesd as the generator network and a four layers convolution network is uesd as the discriminator network.It uses the sparsity of data images to impose texture constraints on the generated images.This work can make the generated images closer to the real images.And then the method makes anomaly judgments based on the prediction errors.It improves the AUC index of the detection models and overcomes the problem of data sparsity to a certain extent.3.Anomaly detection systems for sparse data images are implemented.According to different data requirements,the spatial anomaly detection system and the spatio-temporal anomaly detection system are realized respectively,and the detection results are displayed. |
