Unsupervised Anomaly Detection in Time Series Using LSTM-Based Autoencoders

Automatic anomaly detection in data mining has a wide range of applications such as fraud detection, system health monitoring, fault detection, event detection systems in sensor networks, and so on. The main challenge related to such problem is unknown nature of the anomaly. Therefore, it is impossible to use classical machine learning techniques to train the model, as we don't have labels of time series with anomaly. For periodic time series it is advisable to use STL decomposition of the signal. In such case anomaly detection task is reduced to residuals peak detection. If time series is not periodic (for example, forex price or sound) the only way is using machine learning methods. One of the best machine learning methods is autoencoder-based anomaly detection. An autoencoder is a type of artificial neural network used to learn efficient data encodings in an unsupervised manner. The aim of an autoencoder is to learn a representation (encoding) for a set of data, typically for dimensionality reduction, by training the network to ignore signal “noise”. Unsupervised anomaly detection method based on autoencoders was tested on two types of data: various artificial signal datasets and detection of rare sound events dataset.