Time Series Forecasting and Point Anomaly Detection of Sensor Signals Using LSTM Neural Network Architectures

The sensor data points that exhibit unexpected behaviour that considerably deviates from the norm are considered anomalies in the time-series sensor data. We can model univariate time series data using a number of traditional methods, including ARIMA, GARCH, SARIMA and VAR. Modern deep learning algorithms have lately been used to study time series analysis and prediction. This study contrasts three autoregressive models, which project future events based on past observations. Three different sets of temperature, vibration, and pressure sensor data are used to compare the performance of the Bidirectional Long Short-Term Memory (Bi-LSTM), Convolutional Long Short-Term Memory (C-LSTM), and Stacked Long Short-Term Memory (S-LSTM) architectures. The models performance and training time are reported for healthy, unhealthy and noisy time-series data. Experimental results shows that, Model trained and build using Bi-LSTM can consistently detect point anomaly for healthy, unhealthy and noisy time-series data with minimum error rate across three sensor datasets.