Nowadays, with the rapid growth of data in variation and volume, highlights are on autoanalysis methods. autoanalysis is a technique for extracting information from data. one of the most popular applications for autoanalysis is anomaly detection as a basic machine learning problem. the purpose of anomaly detection is to provide a diagnostic understanding of the abnormal data generative process by assuming that the probability or likelihood of the process generating the normal data is as large as possible. the time series anomaly detection scenarios arise in the context of many applications such as medical data, sensor data or network intrusion. this thesis a new anomaly detection method on time series with an unsupervised approach and by the use of in temporal data the data are not expected to change abruptly unless there are abnormal processes at work.in LSTM explored. the proposed method divided into 2 main parts: multi-step predictions and mods declaration. error predictions from the first part and fit them to three normal distributions in the second part. the likelihood preserving serially dependent values in time series by LSTM made us a powerful tool in both parts. we calculated of each prediction errors suggests as an anomaly score. the results and experiments show that our new approach has a good performance even better than its related works.
