Karimanzira, DivasDivasKarimanziraRitzau, LindaLindaRitzauMartin, TobiasTobiasMartinFischer, ThiloThiloFischer2023-11-212023-11-212023https://publica.fraunhofer.de/handle/publica/45708710.12691/aees-11-3-2It is very important in sensor networks for monitoring, e.g. groundwater quality, to detect sensor failures and anomalous spatio-temporal events such as spills and identify affected areas. Most of the method for anomaly detection do not truly utilize spatial and temporal information. In this paper a novel method based on deep learning (DL) is proposed which truly utilize multivariate spatio-temporal information in anomalous events detection. Anomalous events are quite rare, which makes it very challenging to obtain labeled anomaly datasets. It is therefore purposeful to use an unsupervised approach for sensor anomaly and event detection with labels only being used to set thresholds on prediction errors. Two method for an unsupervised anomaly detection in multivariate spatio-temporal data using deep learning are proposed in this paper. The first framework is composed of a Long Short Term Memory (LSTM) Encoder followed by an LSTM decoder and a LSTM predictor for temporal anomaly detection. In a further step, a Deep Neural Network (DNN) based classifier is used to classify the encoded and trained latent representation to their spatial corresponding to form a temporal and spatial anomaly detector. The second framework is based on a CNN encoder and a LSTM decoder to capture both spatial and temporal features. The encoder component can use either 3D convolutions or Multichannel CNN to capture complex spatial dependencies in each spatial neighborhood. The 3D tensor input for the encoder is formed by stacking the data from the nearest spatial neighbors of each data point. Both methods produce similar results for event detection, detecting different types of anomalies (point, context, etc.). After the training phase to learn the normal system behavior, both methods are capable of detecting anomalies that have never been seen before with a very good accuracy (values ranging between 88% and 96%). To validate the accuracy and efficiency of the DL-based methods, they were compared to a modified ST-DBSCAN algorithm. The results show the superiority of the DL-based methods.enjournal article