Unsupervised anomaly detection with LSTM autoencoders using statistical data-filtering

To address one of the most challenging industry problems, we develop an enhanced training algorithm for anomaly detection in unlabelled sequential data such as time-series. We propose outputs a well-designed system are drawn from unknown probability distribution, U, normal conditions. introduce criterion based on classical central limit theorem that allows evaluation likelihood data-point is U. This enables labelling fly. Non-anomalous passed to train deep Long Short-Term Memory (LSTM) autoencoder distinguishes anomalies when reconstruction error exceeds threshold. illustrate our algorithm’s efficacy, consider two real industrial case studies where gradually-developing and abrupt occur. Moreover, compare performance with four recent widely used algorithms domain. show achieves considerably better results it timely detects while others either miss or lag doing so.