2 research outputs found
Comparative Study of Transformer & LSTM Architectures for Anomaly Detection in Multivariate Time Series Data
Full text linkAnomalydetectionofmultivariatetimeseriesdataiscritical to various industries as it may contain information about the fault in the system or a security breach.Using Long Short-Term Memory(LSTM)networkshaslongbeenthedefault approach for tackling this modeling challenge. Recent Transformer architectures, however, may deliver better performance. The transformer model and the standard LSTM model using RNN architecture are compared through this study to detect anomalies in multivariate time series data. The Jena Climate dataset was used to train both models on the temperature, humidity and air density values to make future predictions and find anomaly from the prediction error. Both models aretrainedtobenext-stepforecasters. Whenthereconstructionerrorisabove an error high threshold, a flag is raised. The LSTM has slightly lower point-wise error than the Transformer, but not necessarily better trends. Most importantly, both designs to identify almost all injected anomalies. The LSTM is more practical for edge deployment because it has far fewer parameters and trains faster. But whenglobal structures play a more important role than resources, Transformers are worth a look.Anomalydetectionofmultivariatetimeseriesdataiscritical to various industries as it may contain information about the fault in the system or a security breach.Using Long Short-Term Memory(LSTM)networkshaslongbeenthedefault approach for tackling this modeling challenge. Recent Transformer architectures, however, may deliver better performance. The transformer model and the standard LSTM model using RNN architecture are compared through this study to detect anomalies in multivariate time series data. The Jena Climate dataset was used to train both models on the temperature, humidity and air density values to make future predictions and find anomaly from the prediction error. Both models aretrainedtobenext-stepforecasters. Whenthereconstructionerrorisabove an error high threshold, a flag is raised. The LSTM has slightly lower point-wise error than the Transformer, but not necessarily better trends. Most importantly, both designs to identify almost all injected anomalies. The LSTM is more practical for edge deployment because it has far fewer parameters and trains faster. But whenglobal structures play a more important role than resources, Transformers are worth a look
