Anomaly Detection on Time Series Sensor Data Using Deep LSTM-Autoencoder

Anomaly detection is crucial in various applications (e.g., cybersecurity, manufacturing, finance, IoT), and an automatic and reliable anomaly detection tool is necessary for accurate prediction. The proposed method in this paper focuses on using deep LSTM Autoencoder on time-series data from loT water level sensors deployed on a water catchment. The method uses unsupervised anomaly detection with deviation methods, which involves lower-dimensional embeddings and reconstruction error. The LSTM Autoencoder model includes feature selection by keeping vital features, and learns the time series' encoded representation. The LSTM model is trained for prediction with three hidden layers based on the encoder's latent layer output. Afterwards, given the output from the prediction model if the reconstruction loss of a data point is greater than reconstruction error threshold the value will be labeled anomaly. We also propose and compare an unconventional method of calculating reconstruction error of each sequence with an aim of reducing false positives and false negatives then compare it with frequently used method. The results show that the LSTM Autoencoder performs well on noisy and real-world datasets for detecting anomalies and also the proposed unconventional method of calculating reconstruction loss increases the models accuracy in identifying anomalies.