INFERENCE OF HELICOPTER AIRFRAME CONDITION Waljinder S. Gill, Ian T. Nabney Nonlinearity and Complexity Research Group School of Engineering and Applied Sciences Aston University, Birmingham, UK Daniel Wells AgustaWestland Ltd. Yeovil, UK ABSTRACT The goal of this paper is to model normal airframe con- ditions for helicopters in order to detect changes. This is done by inferring the flying state using a selection of sen- sors and frequency bands that are best for discriminating be- tween different states. We used non-linear state-space mod- els (NLSSM) for modelling flight conditions based on short- time frequency analysis of the vibration data and embedded the models in a switching framework to detect transitions be- tween states. We then created a density model (using a Gaus- sian mixture model) for the NLSSM innovations: this pro- vides a model for normal operation. To validate our approach, we used data with added synthetic abnormalities which was detected as low-probability periods. The model of normality gave good indications of faults during the flight, in the form of low probabilities under the model, with high accuracy (>92 %). Index Terms— Condition monitoring, vibration, signal pro- cessing, flight condition, switching state space, non-linear model 1. INTRODUCTION The main objective of the project † of which this paper forms a part is to enhance the already effective health-monitoring system (HUMS) for helicopters by analysing structural vi- brational signals to improve further the understanding of air- frame condition. Gearbox monitoring through vibration sig- nal processing is well established [1] but, to our knowledge, there has been no work on the use of vibration data for air- frame structural health monitoring. Past approaches to he- licopter transmission-health monitoring with vibration data have used simple features with direct classifiers and had too many false positives (≥ 30 %) to be practical [2, 3] and there is no prior art for pattern analysis for airframe condition. Thus there is a necessity to develop a more sophisticated approach to achieve a significant advance in predictive maintenance for helicopters, improving safety and reliability at lower cost. † Thanks to EPSRC and AgustaWestland Ltd. for industrial CASE (1000239X) funding. Vibration information during flight is provided by sensors located at different parts of the aircraft. So that structural health can be inferred, features (i.e. sensors and frequency bands) must be chosen which provide the best information on the flight mode of the aircraft. (Compared to fixed wing air- crafts, rotorcraft are characterised by a greater number of dis- tinct flight states such as: steep approach, normal approach, hover, forward flight etc.) These selected features are then used to infer the flight modes and eventually the structural health of the aircraft. The purpose of this paper is to show how flight mode can be inferred accurately from the vibration data and this information can be used to detect abnormalities in the vibration signature. The data provided by AgustaWest- land Ltd. is continuously recorded vibration signals from 8 different sensors during flight. Each sensor measures the vi- bration in a particular direction at a chosen location on the aircraft. During test flights, the aircraft carries out certain planned manoeuvres and we use the knowledge of these ma- noeuvres to help build the models on the labelled data. These models can then be applied to unlabelled data and automati- cally categorise the flight state based purely on sensor mea- surements. The construction of flight-state models from vi- bration data is completely novel; indeed, to our knowledge, there is no prior work on models of different flight modes for helicopters (as opposed to fixed-wing aircraft). 1.1. Methodology Our approach is to build models using features that capture (non-stationary) frequency information by applying a short- time Fourier transform. In this way, it is possible to detect certain signatures or intensities at fundamental frequencies and their higher harmonics. Many of the key frequencies are related to the period of either the main or tail rotor. The in- tensity at these frequencies is greater during certain periods of time and these periods can be associated with flight conditions and transition periods. The frequency resolution we selected yields around 100 features (frequency bands) for each signal. If we were to use all the features from all the sensors together this would give a total of 800 features, which is too high- dimensional for practical modelling and inference. To reduce