U RV QW (sin Θ)g + a x. where. D ΦΘ is determined assuming σ Φ, σ Θ are small and using the approximations

Transcription

1 1 American Control Conference (ACC) Boston Marriott Copley Place July -8, 1. Boston, MA, USA Aircraft Sensor Fault Detection Using State and Input Estimation Ahmad Ansari and Dennis S. Bernstein Abstract This paper presents a method for detecting aircraft faults using state and input estimation. We formulate the kinematics as a nonlinear state space system, which requires no modeling information, and thus is applicable to all aircraft. To illustrate the method, we investigate three fault-detection scenarios, namely, faulty pitot tube, angle-ofattack, and accelerometers. We use the extended Kalman filter for pitot-tube and angle-of-attack fault detection, and retrospective cost input estimation for accelerometer fault detection. For numerical illustration, we use the NASA Generic Transport Model to detect stuck, bias, drift, and deadzone faults. I. INTRODUCTION Sensor failure can have catastrophic consequences for systems that operate under feedback control. The tragic crash of Air France flight 7 on June 1, 9 due to a faulty air-speed is a well-known instance. Techniques for assessing health are thus of intense interest [1 1]. One approach to detecting faults is to use measurements from a set of s to a state or input and then compare the to measurements from a suspect that is not used for estimation. In the case where the suspect corresponds to a state of a dynamic system, state estimation techniques, such as the Kalman filter and its nonlinear variants, can be used. In the case where the suspect corresponds to an input of a dynamic system, state estimation techniques that include input estimation can be used. Input estimation techniques are given in [11 17]. The present paper focuses on fault detection for aircraft flight. In this application, two distinct classes of s are available. Inertial s measure the motion of the vehicle relative to an inertial frame. Relevant s include rate gyros, accelerometers, and vertical gyro. Noninertial s include position measurements from GPS as well as aerodynamic s, such as a pitot tube for measuring forward velocity relative to the air and angle-of-attack (α) and sideslip (β) s for measuring the direction of the relative wind in the body frame. In the present paper we use combinations of inertial and aerodynamic s along with state and input estimation techniques to detect faults. This work is motivated by [8,1], which uses rate-gyro, accelerometer, GPS, angleof-attack, and sideslip measurements to forward velocity relative to the air in order to assess the health of the pitot tube. Ahmad Ansari, and Dennis S. Bernstein are with the Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, USA. {ansahmad, dsbaero}@umich.edu The present paper extends the approach of [8,1] in several ways. First, for pitot-tube fault detection, we apply the extended Kalman filter with augmented bias states in order to deal with biased accelerometer measurements. Unlike [8,1], we do not use GPS to assess the health of the pitot tube. Next, we consider two scenarios that are not considered in [8,1], one of which depends on state estimation and the other on input estimation. In the first scenario, we use the pitot tube, inertial s, and β- to assess the health of the α-, whereas, in the second scenario, we use the pitot tube, rate gyros, vertical gyro, α-, and β- to assess the health of the accelerometers. For input estimation in the second scenario, we use a variation of retrospective cost input estimation as described in [18]. II. PROBLEM FORMULATION A. Aircraft Kinematics The Earth frame and aircraft body-fixed frame are denoted by F E and F AC, respectively. We assume that F E is an inertial frame and the Earth is flat. The origin O E of F E is any convenient point fixed on the Earth. The axes î E and ĵ E are horizontal, while the axis ˆk E points downward. F AC is defined with î AC pointing out the nose of the aircraft, ĵ AC pointing out the right wing, and ˆk AC downward, that is, ˆk AC = î AC ĵ AC. F AC and F E are related by F AC = R AC/E F E, (1) where R AC/E is a physical rotation matrix represented by a 3--1 Euler rotation sequence, involving two intermediate frames F E and F E. In particular, RAC/E = Rî E (Φ) Rĵ E (Θ) (Ψ), () RˆkE where F E = R E /E F E, F E = R E /E F E, and Rˆn(κ) is the Rodrigues rotation about the eigenaxis ˆn through the eigenangle κ according to the right-hand rule. At each time instant, let a denote the air particle located at a point that is fixed relative to the aircraft and upstream of the pitot tube. The location of the aircraft center of mass c relative to O E at each time instant is given by r c/oe = r c/a + r a/oe. (3) Differentiating (3) with respect to F E yields V c = V AC + V a, () where V c = r c/oe, V AC = r c/a, V a = r a/oe. (5) The angular velocity ω AC/E of F AC relative to F E is related to the rotation matrix R AC/E by Poisson s equation /$31. 1 AACC 5951

2 AC R AC/E = R AC/E ω AC/E. () We resolve V AC and ω AC/E in F AC using the notation U V = V P AC, Q = W AC ω AC/E. (7) R AC Resolving the gravity vector g in F AC yields (sin Θ)g g = (sin Φ)(cos Θ)g, (8) AC (cos Φ)(cos Θ)g where g = 3.17ft/s. For the Euler rotation sequence 3--1 (yaw-pitch-roll), we have ω AC/E = AC Φî AC + Θĵ E + Ψˆk E. (9) The acceleration of the aircraft center of mass relative to O E is given by a c/oe/e = v c/oe/e = v c/a/e + v a/oe/e. (1) We assume that the ambient wind is spatially uniform and constant with respect to F E, i.e., v a/oe/e =. Hence a c/oe/e = v c/a/e = Using the transport theorem with (11) yields a c/oe/e = AC V AC. (11) V AC + ω AC/E V AC. (1) Note that (1) is a kinematic relation that is applicable to all aircraft and is independent of all modeling information. B. Kinematic Equations The accelerometer measurement a meas with gravity offset is given by a meas = a c/oe/e g, (13) where the accelerometers are assumed to be located at the center of mass of the aircraft. Substituting (13) into (1) yields AC V AC = ω AC/E V AC + g + a meas. (1) We resolve a meas in F AC using the notation a x a y = a meas. (15) a z AC Resolving (9) in F AC using (7) yields P 1 sin Θ Φ Q = cos Φ (cos Θ) sin Φ Θ. (1) R sin Φ (cos Θ) cos Φ Ψ The inverse transformation of (1) is given by Φ 1 (sin Φ) tan Θ (cos Φ) tan Θ Θ = cos Φ sin Φ P Q. (17) Ψ (sin Φ) sec Θ (cos Φ) sec Θ R Resolving (1) in F AC using (7), (8), and (15) yields U = RV QW (sin Θ)g + a x, (18) V = RU + P W + (sin Φ)(cos Θ)g + a y, (19) Ẇ = QU P V + (cos Φ)(cos Θ)g + a z. () Using the components of V AC resolved in F AC, the angle of attack α and sideslip β are given by α = atan(w, U), β = atan(v, U + W ). (1) Note that (3) (1) are exact kinematic equations, and thus are applicable to all rigid aircraft. Note also that (13) (1) do not include noise. C. Fault-Detection Scenarios Table I lists the available on-board s for fault detection. A continuous-time state-space model can be formulated using (18) (1) as ẋ = f c (x, u k, u u ) + D 1 w, () y = h (x) + D v, (3) where x R lx is the unknown state, u k R lu k is the known input, u u R luu is the unknown input, D 1 w R lx is the process noise with covariance V 1 = D1 D1 T R lx lx, y R ly is the output measurement, and D v R ly is the measurement noise with covariance V = D D T R ly ly, and the functions f c and h are known. The process and measurement noise in () (3) arise due to the fact that the measurements are noisy. Table II defines x, u u, u k, and y for three fault-detection scenarios, in particular, faulty pitot tube, α-, and accelerometers. In each case, the state equations can be written using (18) () and () with l x = 3 as U RV QW (sin Θ)g + a x V = RU + P W + (sin Φ)(cos Θ)g + a y +D 1 w. () Ẇ QU P V + (cos Φ)(cos Θ)g + a z In the absence of vertical gyro measurements of Φ and Θ, the first two equations in (17) can be integrated to obtain s of Φ and Θ. The matrices D 1 and D are determined using Tables I II and are given as follows; Scenario 1) Faulty pitot tube: D 1 = [ ] D P QR D ΦΘ D axa ya z, D = diag (σ α, σ β ), where (5) W V σ P D P QR = W U σ Q, V U σ R cos Φ [ ] D ΦΘ = (sin Φ) sin Θ (cos Φ) cos Θ σφ, σ (cos Φ) sin Θ (sin Φ) cos Θ Θ σ ax D axa ya z = σ ay. σ az D ΦΘ is determined assuming σ Φ, σ Θ are small and using the approximations sin w Φ w Φ, sin w Θ w Θ, cos w Φ 1, cos w Θ 1, w Φ w Θ = w Θ w Φ, where w Φ and w Θ are noise on the measurements or s of Φ and Θ, respectively, obtained from (17). Scenario ) Faulty α-: D 1 is given by (5) and D = diag (σ U, σ β ). () Scenario 3) Faulty accelerometer: D 1 = [ D P QR D ΦΘ ], D = diag (σ U, σ α, σ β ). (7) 595

3 TABLE I: On-board s for fault detection. The additive noise for each is assumed to be white Gaussian. Sensors Measurements Standard Deviation of Noise Pitot Tube U σ U Rate Gyro P, Q, R σ P, σ Q, σ R Vertical Gyro Θ, Φ σ Θ,σ Φ Accelerometers a x, a y, a z σ ax, σ ay, σ az α- α σ α β- β σ β TABLE II: Scenarios for fault detection Faulty Sensor x y u k u u a x, a y, a z Pitot Tube U, V, W α, β P, Q, R Θ, Φ a x, a y, a z α- U, V, W U, β P, Q, R Θ, Φ Accelero- P, Q, R U, V, W α, β, U meter Θ, Φ a x, a y, a z III. STATE AND INPUT ESTIMATION In Section II-C, fault detection is formulated as a problem of state and input estimation. We use the extended Kalman filter (EKF) for state estimation and retrospective cost input estimation (RCIE) for input estimation. The state-space model () (3) can be discretized to first order as x(k + 1) = f (x(k), u k (k), u u (k)) + ˆD 1 (k)w(k), (8) y(k) = h (x(k)) + D v(k), (9) where k is the time step, D1 (k) = T s D 1, f (x(k), u k (k), u u (k)) = x(k) + T s f c (x(k), u k (k), u u (k)), and T s is the sampling time. A. Extended Kalman Filter (EKF) EKF consists of two steps. Assuming that u = u k, that is, the input is fully known, the forecast step is given by x f (k) = f (x da (k 1), u k (k 1)), (3) P f (k) = A(k 1)P da (k 1)A T (k 1) + V 1 (k 1), (31) where x f (k) R lx is the forecast state, x da (k) R lx is the data assimilation state, P f (k) R lx lx is the forecast error covariance, P da (k) R lx lx is the data assimilation error covariance, V1 = D T 1 D1 is the process noise covariance, and A(k) is the Jacobian of f given by A(k) = f x (3) xda (k),u k (k). The data assimilation step is given by K da (k) = P f (k)c T (k)s 1 da (k), (33) P da (k) = P f (k) P f (k)c T (k)s 1 da (k)c(k)p f(k), (3) x da (k) = x f (k) + K da (k) [y(k) h (x f (k))], (35) where K da (k) R lx ly is the state estimator gain, S da (k) = C(k)P f (k)c T (k) + V (k), and C(k) is the Jacobian of h C(k) = h x (3) xf (k). B. Retrospective Cost Input Estimation In the case where the input is partially or fully unknown, (3) does not explicitly account for the unknown input u u. The effect of u u (k) can be included in the process noise w(k) by a suitable choice of Q(k). A more effective approach is to u u (k) and include it in (3) and (3) with its û u (k) as x f (k) = f (x da (k 1), u k (k 1), û u (k 1)), (37) A(k) = f x (38) xda (k),u k (k),û u(k). In order to û u (k), we construct an adaptive input estimator that minimizes z(k) = y(k) h (x da (k)), (39) where z(k) R lz is the output error. The d input û u (k) is the output of the input estimation subsystem of order n c given by n c n c û u (k) = Mi T (k)û T u (k i) + Ni T (k)z T (k i), i=1 i=1 () where M i (k) R l ûu l ûu and N i (k) R ly l ûu. The subsystem in () can be reformulated as where l θ = lûu n c (lûu + l y ), and û u (k) = Φ(k)θ(k), (1) θ(k) = vec([m T 1 (k) M T n c (k) N T 1 (k) N T n c (k)] T ) R l θ, φ(k) = [û T u (k 1) û T u (k n c ) y T (k 1) y T (k n c )] T, Φ(k) = I lûu φ T (k) R l ûu l θ, () where denotes the Kronecker product, Φ(k) is the regressor, and vec is the column-stacking operator. 1) Retrospective Performance: Define G f (q) = D 1 f (q)n f (q), where q is the forward shift operator, n f 1 is the order of G f, and N f (q) = K 1 q n f 1 + K q n f + + K nf, (3) D f (q) = I lz q n f + A 1 q n f 1 + A q n f + + A nf. () Furthermore, K i R lz l ûu for 1 i r, Aj R lz lz for 1 j r, and det (D f (q)) is asymptotically stable. Next, for k 1, we define the retrospective performance variable ẑ(ˆθ, k) = z(k) + Φ f (k)ˆθ u f (k), (5) where Φ f (k) = G f (q)φ(k), u f (k) = G f (q)û u (k), () and ˆθ R l θ is determined by optimization below. ) Markov Parameters: The filter G f at time step k is based on the input-to-performance transfer matrix G zuu (q, k) = C(k)(qI A(k)) 1 B(k), where A(k) is defined by (38), C(k) is defined by (3), and B(k) is the 5953

4 Jacobian of f with respect to û u evaluated at û u (k 1). For all complex numbers z whose absolute value is greater than the spectral radius of A(k), it follows that H i (k) G zuu (z, k) = z i, (7) i= where, for all, i 1, the i th Markov parameter of G zuu (z, k) is defined by H i (k) = C(k)A(k) i 1 B(k). (8) A. Types of faults We consider the following types of faults: Bias. The measurement has a constant offset from the measurement. Drift. The measurement has a constant-slope deviation from the measurement. Deadzone. The reads zero within a specific range. Stuck. The reading is fixed. B. Procedure for fault detection For fault detection using state and input estimation, rich signals are needed. This can be achieved either by exciting the dynamics of the aircraft using its control surfaces or it can arise naturally from the atmosphere, e.g., wind gusts. For this paper, the dynamics of the aircraft are excited using a saturated harmonic elevator input. For fault detection, we assume the remaining s are functional, and we define the residual e(k) = k+kw [y(i) y est (i)], (51) i=k where y est is the of y, and k w is the data-window size. We call e as for value of y, and for measured value of y. For each numerical example, k w = 1 data points, and x da () =.5x(). At each time step k, G f is chosen to be the finite-impulseresponse (FIR) filter n f H i (k) G f (q, k) = q i, (9) V. GTM EXAMPLES i= We set the sampling time T s =.1 s and consider a obtained by truncating (7). scenario where GTM is initially trimmed for level flight at an 3) Cumulative Cost and RCIE Update Law: For k 1, altitude of 8 ft. We excite the aircraft dynamics using the we define the cumulative cost function elevator deflection δe(k) = sat [ sin(kt s + 5)] deg, J(k, ˆθ) k ( ) = ẑ(i) T R z ẑ(i) + [Φ(i)ˆθ] T which is a saturated sinusoid with amplitude deg, maximum deflection of ± deg, and a period of s. Unless stated Rûu Φ(i)ˆθ i=1 otherwise, the ambient wind is constant with magnitude of + [ˆθ θ()] T R θ [ˆθ θ()], (5) 1.88 ft/s. where R θ, R z, and Rûu are positive definite. Let P () = To emulate noise, we add zero-mean white noise R 1 θ and θ() = θ. Then, for all k 1, the cumulative to all of the measurements with standard deviations cost function (5) has the unique global minimizer ˆθ = θ(k) σ ax = σ ay = σ az =.1g, σ P = σ Q = σ R =.1 rad/s, given by the RLS update σ Φ = σ Θ =.1 rad, σ α = σ β =.1 rad, and σ U = θ(k) = θ(k 1) P (k 1) Φ(k) T Γ(k) 1 [ Φ(k)θ(k 1) + z(k)], 1. ft/s. To show the fault and noise level, we plot the measurement and measurement together. We also present the where P (k) satisfies P (k) = P (k 1) P (k 1) Φ(k) T Γ(k) 1 Φ(k)P (k 1), to show the accuracy of the s. Note that [ ] in practical application, the measurement and thus the Φ(k) = Φf (k) R (lz+l ûu ) l θ, are not available. However, the Φ(k) [ ] can be used in practice for fault detection. R(k) = Rz (k) R (lz+l ûu ) (lz+l ûu ), Rûu (k) A. Fault Detection for Pitot-Tube Failure [ ] In the following cases, the pitot tube fails by becoming z(k) = z(k) uf (k) R lz+l ûu, stuck at the constant value of 1 ft/s, beginning at 1 s. First we U with noisy input and output measurements. Fig. 1 shows that the decreases to Γ(k) = R(k) 1 + Φ(k)P (k 1) Φ(k) T. IV. SENSOR FAULT DETECTION ft/s, indicating that EKF is operating correctly. However, the The formulation in sections II III is applicable to all rigid jumps after the fails. aircraft. In this paper, we use the NASA Generic Transport Next, to see the effect of accelerometer bias, we consider Model (GTM) to illustrate fault detection. GTM is a the case with accelerometer bias b ax = b ay = b az =.1g. high-fidelity six-degree-of-freedom nonlinear aircraft model Fig. shows that the d pitot-tube measurement drifts with aerodynamic lookup tables [19]. due to the biased accelerometers. In order to deal with these biases, the dynamics in () are augmented as 595 ẋ = f c (x, u k, u u ) + ˆb + D 1 w, ˆb =, (5) where ˆb R 3 is the d bias in the accelerometers. Fig. 3 shows that, with the augmented states, the is less than ft/s, thus indicating no drift in the of U. Consequently, ˆb converges to the accelerometer bias. B. Fault Detection for α- Failure We now present cases where the α- has either a bias or deadzone beginning at t = 1 s. We α using ˆα = atan(ŵ, Û), (53) where Ŵ and Û are the state s of (). First, we consider the case where the α- has a bias

5 U (ft/s) e (ft/s) α (deg) e (deg) Fig. 1: Stuck pitot tube. At 1 s, the measurement is stuck at 1 ft/s. The jumps to a mean value of 8.5 ft/s indicating pitot-tube failure. U (ft/s) e (ft/s) Fig. : Estimation of U with biased accelerometers. The of U drifts from the measurement. Beginning at 1 s, the measurement is stuck at 1 ft/s. The and s are both increasing, and therefore it is not possible to detect the fault. This shortcoming is overcome in Fig. 3. U (ft/s) e (ft/s) Fig. 3: Estimation of U with augmented bias states. The of U indicates no drift. Beginning at 1 s, the measurement is stuck at 1 ft/s. The is less than ft/s, whereas the has an offset due to the stuck fault. of deg. Fig. shows that the is less than Fig. : α- with a bias. Beginning at 1 s, the α- has a bias of deg. The indicates an offset due to the bias. α (deg) e (deg) Fig. 5: α- with a deadzone. Beginning at 1 s, the α- reads zero within ± deg. The indicates an offset due to the deadzone. deg, and the jumps to.5 deg due to the bias. Next, we consider the deadzone case where the α- reads zero within ± deg. Fig. 5 shows that the residual has an offset due to the deadzone. C. Fault Detection for Accelerometer Failure For accelerometer fault detection, we use RCIE to acceleration. To do so, we filter U and augment the states in () with an additional filtered state U f as U f = ω c (U f U), (5) where ω c > is the cutoff frequency. We use U f as the output measurement. We choose ω c = π rad/s, n c =, n f =, R θ = 1 I lθ, and R u = 1 3. Furthermore, we choose R z = 1 and R z = 1 to a x and a z, respectively. We consider cases where the accelerometer has either a bias or drift beginning at 1 s. Fig. shows that RCIE is able to a x and a z. Figs. 7a and 8a show cases where a x and a z have biases of.5g and.1g, respectively. Note that the s have offsets due to these biases. Figs. 7b and 8b show cases where the measurements of a x and a z drift with a slope of.1 g/s. Note that the residuals also drift from the respective s. 5955

6 a x (g) d actual a z (g) d actual Fig. : Acceleration estimation using RCIE. Note that, RCIE is able to a x and a z. Estimation of a x. Estimation of a z Fig. 7: a x-. The measurement of a x is subject to a bias. Note that the jumps at 1 s when the bias begins. The measurement of a x is subject to a drift. Note that the begins to increase at 1 s when the drift begins Fig. 8: a z-. The measurement of a z is subject to a bias. Note that the jumps at 1 s when the bias begins. The measurement of a z is subject to a drift. Note that the begins to increase at 1 s when the drift begins. VI. CONCLUSIONS In this paper, we presented a method for detecting aircraft faults using state and input estimation. We used the extended Kalman filter (EKF) and retrospective cost input estimation (RCIE) to states and inputs, respectively. For the estimation framework, we used the kinematics to formulate a nonlinear state space system. Three fault-detection scenarios, in particular, faulty pitot tube, angle-of-attack, and accelerometers were investigated. We used EKF for pitot tube and angle-of-attack fault detection, and RCIE for accelerometer fault detection. In order to illustrate fault detection, we used the NASA Generic Transport Model and presented cases for detecting stuck, bias, drift, and deadzone faults. For all cases, we showed that the can be used to detect faults. VII. ACKNOWLEDGMENT This research was supported in part by the Office of Naval Research under grant N [1] R. Isermann, Process fault detection based on modeling and estimation methods-a survey, Automatica, vol., pp. 387, 198. [] A. M. Agogino, S. Srinivas, and K. M. Schneider, Multiple expert system for diagnostic reasoning, monitoring and control of mechanical systems, Mech. Sys. Sig. Proc., vol., pp , [3] D. Zhou and P. Frank, Fault diagnostics and fault tolerant control, IEEE Trans. Aerosp. Elec. Sys., vol. 3, pp. 7, [] J. Gertler, Fault Detection and Diagnosis in Engineering Systems. CRC press, [5] R. Isermann, Fault-Diagnosis Systems: An Introduction from Fault Detection to Fault Tolerance. Springer,. [] Y. Zhang and J. Jiang, Bibliographical review on reconfigurable faulttolerant control systems, Ann. Rev. Contr., vol. 3, pp. 9 5, 8. [7] P. Freeman, P. Seiler, and G. J. Balas, Air data system fault modeling and detection, Contr. Engr. Pract., vol. 1, pp , 13. [8] S. Gururajan, M. Fravolini, M. Rhudy, A. Moschitta, and M. Napolitano, Evaluation of failure detection, identification and accommodation (SFDIA) performance following common-mode failures of pitot tubes, SAE Technical Paper, Sept 1. [9] M. L. Fravolini, M. Rhudy, S. Gururajan, S. Cascianelli, and M. Napolitano, Experimental evaluation of two pitot free analytical redundancy techniques for the estimation of the airspeed of an uav, SAE Int. Jr. of Aerosp., vol. 7, pp , 1. [1] M. B. Rhudy, M. L. Fravolini, Y. Gu, M. R. Napolitano, S. Gururajan, and H. Chao, Aircraft model-independent airspeed estimation without pitot tube measurements, Aerosp. and Elec. Sys., IEEE Transactions on, vol. 51, pp , 15. [11] M. Corless and J. Tu, State and input estimation for a class of uncertain systems, Automatica, vol. 3, pp , [1] Y. Xiong and M. Saif, Unknown disturbance inputs estimation based on a state functional observer design, Automatica, vol. 39, pp , 3. [13] T. Floquet and J. P. Barbot, State and unknown input estimation for linear discrete-time systems, Automa., vol., pp ,. [1] S. Gillijns and B. De Moor, Unbiased minimum-variance input and state estimation for linear discrete-time systems, Automatica, vol. 3, pp , 7. [15] H. Palanthandalam-Madapusi and D. S. Bernstein, A subspace algorithm for simultaneous identification and input reconstruction, Int. J. Adaptive Contr. Sig. Proc., vol. 3, pp , 9. [1] S. Kirtikar, H. Palanthandalam-Madapusi, E. Zattoni, and D. S. Bernstein, Delay input and initial-state reconstruction for discrete-time linear systems, Circ. Sys. Sig. Processing, vol. 3, pp. 33, 11. [17] H. Fang, R. A. de Callafon, and J. Cortes, Simultaneous input and state estimation for nonlinear systems with applications to flow field estimation, Automatica, vol. 9, pp , 13. [18] A. M. D Amato and D. S. Bernstein, Adaptive forward-propagating input reconstruction for nonminimum-phase systems, Proc. Amer. Conf. Contr., pp , Montreal, 1. [19] T. Jordan, W. Langford, C. Belcastro, J. Foster, G. Shah, G. Howland, and R. Kidd, Development of a dynamically scaled generic transport model testbed for flight research experiments, AUVSI Unmanned Unlimited, Arlington, VA,. 595