A multivariate statistical approach for anomaly detection and condition based maintenance in complex systems

Abstract

This paper will describe recent developments in the practical application of multivariate Statistical Process Control (SPC) techniques to improve and automate anomaly detection in complex systems as part of a comprehensive Condition Based Maintenance (CBM) system. The CBM implementation process will be broken down into phases from planning to operation. The operational CBM system includes data acquisition, preprocessing, detection, and diagnosis. The multivariate statistical methods for establishing baseline operation and analyzing operational data are covered. A comprehensive discussion of the technical need, complexities of implementation, and some best practices based upon novel applications of the technology are included. This technology is applicable to complex mechanical systems exhibiting degradation or wear prior to system failure and where failure leads to significant financial losses or loss of mission capability. Examples of the implementation of the methodology will be drawn from jet engine trending, facility infrastructure, utility-scale wind turbines, and projects where the techniques have been successfully employed in real world applications. CBM has been shown to be an effective approach to reducing the Life Cycle Cost of a system and improving system availability. Univariate SPC techniques are widely used in industry but are sometimes insufficient for monitoring complex systems which function over a wide range of operating conditions and environments. Multivariate SPC techniques provide high sensitivity and low false alarm rate monitoring opportunities. Simplifying the system health assessment of a complex system into a single control chart reduces the burden of interpreting complex interactions and reduces the training and education level of the maintenance personnel. Multivariate SPC techniques can also be implemented to improve anomaly detection in environments where it is impossible or impractical for cost or performance reasons to employ additional sensors for fault detection. Significant cost savings and cost avoidance can be realized by utilizing the full life of equipment but also having early detection of impending problems when there is still time for failure mitigation or scheduling of preventative maintenance. The statistical techniques are also useful for forensic analysis of failed systems where multivariate SPC have not been implemented. The authors introduced this technique at the 2016 International Symposium on Systems Engineering in Edinburgh, Scotland [1]. This paper will expand and build upon the previous work by adding the results of recent applications and elaborating on the approach as it pertains to the current state-of-the-art based on these results.