Machine Learning Techniques for Automatic Sensor Fault Detection in HUMS Systems
Refereed Conference Meeting Proceeding
In this paper we describe the problem of developing sensor fault detection within HUMS instrumentation systems, and solutions based upon machine-learning techniques. We conclude with a report on our proof-ofconcept demonstrator, and outline next-steps towards implementation of a autonomous selfdiagnostic sensor solution. 1. INTRODUCTION Good data is key to the success of a health and usage monitoring (HUMS) program, and modem data acquisition systems allow for reliable, high fidelity data capture. Unfortunately HUMS programs are often hindered by undetected sensor and wiring problems that can lead to invalid data and inconclusive analysis. Many authors have identified sensors and wiring as the weakest link in an entire HUMS system (e.g. , ), where the transducer and consistency of the transducer/structure interface can “make or break” a system. Choosing long-life sensors with lifetimes in excess of assets under test is one approach for addressing this problem, however these high cost and high specification sensors are rarely economically viable. Traditionally for airborne health monitoring programs, there is a realistic expectation that sensors will be replaced over time, and that dedicated data analysts will be available to spot subtle signs within data which indicate the onset of sensor/wiring faults. This approach does not scale well for large fleet deployments and does not allow for robust automation. In this paper, we present new sensor diagnostic approaches based on “Machine Learning” techniques1 . These automated techniques allow for reliable measurement using practical cost sensors, installed in extended duration monitoring programs spanning many years. Machine Learning fault detection techniques not only detect the obvious catastrophic sensor errors which clearly manifest themselves in captured data, but also the more subtle sensor issues that can easily go undetected for long periods of time, leading to less reliable structural analysis, e.g. • strain gauge de-lamination • accelerometer de-calibration • dry solder joints (Fig. 1 left) • loose wiring (Fig. 1 right) A key enabler for airborne HUMS systems is the automation of human data analysis, allowing systems to operate reliably without intervention for many years. This paper explores how “Machine Learning” techniques can be used to detect subtle signs of sensor/wiring faults within captured data, essentially automating the experience of human analysts who must ensure captured data is good. Machine Learning techniques offer the ability to not just automatically track condition indicators, but to also learn what the optimal condition indicators are for a particular set of data driven features.
17th Australian International Aerospace Congress: AIAC 2017
Digital Object Identifer (DOI):
National University of Ireland, Dublin (UCD)
Open access repository: