Air Data Virtual Sensor: a Data-Driven Approach to Identify Flight Test Data Suitable for the Learning Process

Governments and main stakeholders from all over the world will make available huge funds to develop a greener aviation. To this aim, important updates are expected in the next years in aerodynamics, A/C configuration, propulsion and onboard systems. In addition, the next advent of the UAVs civil operability, and possible complexity deriving from high level of redundancy, is pushing the aerospace community towards the use of new technologies for a smarter A/C system integration. As far as avionics is concerned, the trend shows that the new avionic paradigms, e.g. Fly byWire and distributed avionics that are successfully applied on large passenger aircraft (e.g. Airbus A380), will be commonly used even on smaller aircraft. The digital revolution experienced in last decades will be crucial to achieve a smarter integration of onboard systems. Air Data Systems will be updated, the most are still based on pneumatic probes or vanes, in order to enable beneficial avionic integration. In recent years, several studies were conducted for a smarter sensor fusion to be used to provide alternate sources of air data with the aim to detect ADS faults avoiding common modes and to provide analytical redundancy. The present work is part of the Smart-ADAHRS project that is born aiming to design a simplex complete air data system partially based on virtual sensors. The main objective of the aforementioned project is to provide an innovative ADS with a lighter configuration (some sensors are replaced by virtual ones) assuring the same performance and reliability of commons ADS. At the moment, the authors are involved to correlate flight test, obtained with a flying demonstrator on an ULM aircraft, and simulated environment performance. The virtual sensors are based on neural network techniques and, therefore, the learning process is crucial to obtain suitable performance. Moreover, using real flight data introduced new uncertainties to the training data set that required a pre-processing of the training data. The present work describes the approach used to extract quasi-steady and quasi-symmetric data from the entire flight data record. The main objective of the tool is to avoid common issues in MLP training (e.g. local minima) and to promote a more uniform distribution of the training data set inside the n-dimension domain where the neural networks are defined.