A Machine Learning Approach to Pitot Static Error Detection & Airspeed Prediction Renee Swischuk 1 , Douglas Allaire 2 1 Texas A&M University, Department of Mathematics 2 Texas A&M University, Department of Mechanical Engineering Objective Error Detection • Two running estimates of airspeed from pitot tube and accelerometer data: () = 2( ()− ()) ρ () = −1 • Anytime | − | > ℎℎ , classify the state of the system. System states: 1. Pitot tube block – Airspeed indicator drops to 0 2. Pitot drain block – Total pressure constant 3. Static port block – Static pressure constant • Hundreds of redundant data streams are recorded in flight. • Sammons Mapping Function 3 was minimized to find lower dimensional data sets which preserve the structure and distances of the original high dimensional data: E= < [ ∗ − ] 2 ∗ < ∗ Classifying the State of the System • Erroneous pressure streams are manually simulated for 100 seconds. • Autocorrelation for each pressure stream is stored in the error library. • Autocorrelation is calculated online every 10 seconds. • Online case is classified into the error library to detect constant signals. • System state is determined using minimum Euclidean distance. Offline Airspeed Library Results Airspeed Prediction Drain Ram Air Static Air Pitot Tube Pitot Connection Diaphragm Handstaff Pinion Long Lever Sector The Effects of Pitot+Drain Block on Total Pressure 47450 47400 47350 47300 47250 47200 47150 47100 Total Pressure (Pa) 0 50 100 150 200 Airspeed Prediction 450 400 350 300 250 200 150 100 50 0 2000 4000 6000 8000 10000 Airspeed (knots) Time (sec) • Airspeed predicted using K-nearest neighbors regression. • This system is able to detect error within 20 seconds. • Airspeed prediction is within 30 knots of the true airspeed during cruise. • This system will allow pilots to safely navigate an aircraft in the event of a pitot static error and provide UAV’s the ability to continue flight. Time (sec) Unmanned Aerial Vehicles only have one pitot static system used for guidance so system failure can result in critical missions being aborted. If autonomous error detection software is developed, these missions may be able to continue on. Motivation Our goal is to create a system capable of detecting pitot static blocks within 30 seconds of failure. Failures result in inaccurate airspeed readings so in addition, we attempt to predict a corrected airspeed which lies within 10 knots of the true airspeed. Approach 1 U.S. Dept. of Transportation, Federal Aviation Administration , Pilot's handbook of aeronautical knowledge, 2008. 2 King, A. D., B.Sc, F.R.I.N., ”Inertial Navigation - Forty Years of Evolution”, GE Company Review, Vol 13, No 3, 1998 3 LLC, CH ROBOTICS. "Understanding Euler Angles.“ November 26th, 2012 4 Camastra, Francesco. "Data dimensionality estimation methods: a survey." Pattern recognition 36.12 (2003): 2945-2954. Inertial Navigation System 1 • In the reduced dimension, the following most informative features for predicting airspeed were chosen using simulated annealing optimization: Body Coordinate System 2 Using highly reliable Inertial Navigation System (INS) data, we can detect pitot static errors and trigger a classification model for identifying the state of the system. In the event of a failure, a corrected airspeed is predicted using k-nearest neighbors in an offline/online paradigm. Initial feature selection and reduction is performed on the high dimensional flight data. References 500 400 300 200 100 0 500 Error (knots) Orion UAV, Aurora Flight Sciences Pitot Static System Diagram 1 0 2000 4000 6000 8000 10000 Time (sec) 500 400 300 200 100 0 ? Example of K-Nearest Neighbor Classification K=3 Class ( ) = ? Choosing K Percent Correct .978 .977 .976 .975 .974 .973 .972 0 2 4 6 8 10 12 14 • Fuel Flow • Engine Core Speed • Engine Fan Speed • Pitch • Angle of Attack • Thrust Target • Thrust Command Roll Pitch Yaw Z Y X Accelerometers Gyros V INS V pitot V predicted Error V thresh System state This work was partially supported by FA9550-16-1-0108, under the Dynamic Data-Driven Application Systems Program, Program Manager Dr. Frederica Darema. 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Number of Data Streams (Dimension) Error Detection and Airspeed Prediction Dimensionality Reduction .040 .035 .030 .025 .020 .015 .010 .005 0.00 -.005 Average Structure Loss Higher dimensionality has minimal gain Number of Neighbors (k) Airspeed (knots)