Anomaly detection using Dynamical Linear Models and sequential testing on a marine engine system

Erik Vanem and Geir Olve Storvik
Submission Type: 
Full Paper
AttachmentSizeTimestamp
phmc_17_021.pdf643.14 KBSeptember 6, 2017 - 10:15am

This paper presents a study on the use of Dynamical Linear Models for anomaly detection and condition monitoring of a marine engine system. Various sensors are installed at different places within the engine system and records essential parameters such as power output from the engine, engine speed, bearing temperatures and various other temperatures, speeds and pressures for selected engine components. The idea is to utilize the information in these sensor signals in order to monitor the condition of the engine. Such a condition monitoring system includes fault detection, diagnosis and prognostics, and robust anomaly detection is a prerequisite for reliable management of the system. Dynamical Linear Models (DLM) constitute a flexible framework for modelling of such sensor signals, where the sensor signals are modelled conditional on some latent states, and the model provides k-step ahead forecasts of the sensor signals that can be compared to new sensor readings. The idea is to establish a model that represents normal operations of the system. Statistical sequential testing will then be performed on the residuals and model breakdown can be an indication of deviation from normal conditions and possible impending failures of the engine system. This will then call for further diagnostics and prognostics tasks to interpret the nature of the deviation, and . The Dynamical Linear Model framework is very flexible and can accommodate a range of candidate models. However, very complicated models in high dimensions may be computationally expensive to estimate and apply, so various pre-processing techniques are investigated to improve model performance, including simple regression models, cluster analysis and principal component transformation. Model selection among the candidate models are based on root mean square errors and the Akaike Information Criterion.

The full paper will include a description of the engine system and the various sensor signals and will specify and estimate model parameters for a few model alternatives. The performance of these models on the sensor data will be reported and supported by figures and tables, as necessary. Moreover, some pre-processing steps applied in order to improve model performance will be outlined and the effect of this on the performance of the models will be reported.

Publication Year: 
2017
Publication Volume: 
8
Publication Control Number: 
021
Page Count: 
16
Submission Keywords: 
Condition monitoring; Time series analysis; Marine applications; Fault detection; Data driven models
Submission Topic Areas: 
Data-driven methods for fault detection, diagnosis, and prognosis
Industrial applications
Sensors
Submitted by: 
  
 
 
 

follow us

PHM Society on Facebook Follow PHM Society on Twitter PHM Society on LinkedIn PHM Society RSS News Feed