Dynamical Linear Models for Condition Monitoring with Multivariate Sensor Data
Geir Olve Storvik
Condition monitoring, anomaly detection, dynamical linear models, ship machinery systems, Bayesian time-series, sensor data
This paper presents an application of dynamical linear models for anomaly detection and condition monitoring of ship machinery systems based on multivariate sensor data. Various model alternatives are specified and fitted to a set of training data before they are applied to a test set. Sequential monitoring based on statistical tests are applied to detect model breakdown as an indication of deviation from normal conditions. The framework is very flexible and allows for a range of different candidate models to be specified. In this paper, some of the estimated models perform rather poorly, but the best ones do quite well in flagging anomalies in the data streams. Hence, it is demonstrated that dynamical linear models may be utilized for anomaly detection and condition monitoring with multivariate sensor data. However, identification of the best model structure is challenging and requires representative training data and careful consideration in the model specification.
Author(s) by submitting the manuscript to the International Journal of COMADEM agree to transfer the rights to COMADEM International UK with some exceptions as described in COPYRIGHT TRANSFER AGREEMENT available here.