Abstract. Nomenclature

Transcription

1 OPEN DESIGN FOR HELICOPTER ACTIVE CONTROL SYSTEMS Geoffrey J. Jeram School of Aerospace Engineering Georgia Institute of Technology Atlanta, Georgia Abstract With a holistic approach, this open design for helicopter active control systems applies intelligent control techniques to provide the pilot with useful and intuitive tactile cues. The design has five interdependent functional modules that provide: limit prediction and avoidance cues based on neural networks; heuristic cues based in fuzzy inference systems; an intelligent arbitrator to distribute multiple and possibly conflicting cues for multiple control axes; and a tactile interface to define appropriate force characteristics for various kinds of cues. Each module is explained as a process in the context of practical applications that include maneuver envelope limit avoidance cues, emergency procedure prompts, instrument cues, robot assistance for routine tasks, and system customization to the pilot. The design is demonstrated as a limit avoidance cueing system in the Real-Time Interactive Prototype Technology Integration Development Environment (RIPTIDE) at the Army/NASA Rotorcraft Division. Nomenclature f, g, h = Vector functions F u x y y p = Vector of active control axes counter forces from the active inceptor to the pilot, F = [ F col F long F lat F pedl ] T = Control vector of k elements u = [ δ 1 δ 2 δ k ] T = Aircraft State Vector of n elements x = [ x 1 x 2 x 3 x n ] T = Aircraft Limit Vector of m elements y = [ y 1 y 2 y 3 y m ] T = Predicted Limit Vector u crit = Control margin vector y = Limit margin vector Subscripts ANN = Adaptive Neural Network coll = Collective long = Longitudinal (cyclic) lat = Lateral (cyclic) crit = Critical fut = Future p = Predicted NN = Neural Network (implies static) * = Assumed (chosen) future time history (as in u * ) Introduction Background and Context The tactile connection between man and machine is a physically direct connection. The aerodynamic forces on the aircraft control surfaces travel through rods and cables to the pilot s arms and legs. These control forces inform the pilot about the aircraft through his muscles and joints. It is a proprioceptive perception, that is, a sense originating within the organism, and it is distinguished from visual and vestibular perception. The pilot s use of this proprioceptive information is commonly referred to as seat of the pants flying. As aircraft grow larger and incorporate hydraulics, power assist servos, and fly-by-wire technology, their control systems sever the physical connection between the pilot s cockpit controls and the aerodynamic forces. The pilot loses a channel of information and it is more difficult for him to get the feel for the aircraft and use it a natural extension of his body. Active control inceptors are cockpit controls such as collective levers, sidesticks, yokes, and cyclic sticks that generate artificial forces against the pilot s hands and feet. With active inceptors, a control system can use forces to provide the pilot with information about the aircraft. Advanced aircraft with highly unconventional configurations operating across several regimes of flight and with highly non-linear control dynamics do not react the same way the conventional airplanes did. But with intelligent

2 limit prediction systems, artificial force cues can return to the pilot a somatic understanding of the aircraft. Moreover, an active inceptor can physically manipulate cockpit controls. An intelligent active control system can serve the pilot as a robot assistant that performs routine tasks, prompts emergency procedures, and otherwise lightens pilot workload. Aircraft limit prediction systems linked to tactile cues in flight controls enable more carefree handling and help us make the most of an aircraft's flight envelope 1. Recent studies such as the Helicopter Maneuver Envelope Enhancement (HELMEE) program have shown as much in the NASA Ames' Vertical Motion Simulator 2. Other ongoing projects such as the Helicopter Active Control Technology program sponsored by the U.S. Army and carried out by Boeing continue to explore the potential of active cueing 3. The Army/NASA Rotorcraft Division continues to develop its Rotorcraft Aircrew Systems Concepts Airborne Laboratory (RASCAL) in a JUH-60 Blackhawk airframe 4. With a two-axis active sidestick controller and a full-authority fly-by-wire flight control system, the RASCAL facilitates active control and limit cueing research. The Army/NASA Rotorcraft Division Flight Mechanics and Cockpit Integration Branch and the School of Aerospace Engineering at the Georgia Institute of Technology are developing the control system for the RASCAL active sidestick. One product of the endeavor is this holistic approach and open design architecture. It is intended for the RASCAL active control system and is applicable to general haptic applications. Motivation and Significance The active control projects of the last decade, including many limit avoidance and tactile cueing projects, have advanced the theory of limit prediction and tactile cueing for avoidance. When implemented in simulation and practical evaluation, the active control systems have been specific designs for targeted applications, as in the HELMEE project, and or are tightly integrated into the aircraft flight control system, as in the Helicopter Active Control Technology (HACT) program 5. This project likewise began as a limit avoidance system targeted for the RASCAL Helicopter, but it took a holistic approach and the design has broadened into an open engineering system. It has an architecture that maintains design freedom, flexibility, and growth potential as far as possible. As limit prediction models are refined and new ones devised, they can be fitted into this architecture, replacing previous algorithms or adding to their capability and accuracy. It is intended for diverse applications in the aerospace industry and beyond. Several specific new concepts and techniques are developed within this overall design. Previous limit avoidance cueing methods successfully mapped a single limit to a single control axis. But these programs treated the limit-to-control relationship as one-to-one. The limit surface search algorithm created with this design uses the non-linear mapping of the limit surface to search across admissible control positions to find the critical position that will cause the aircraft to exceed a limit. Unlike previous methods, it accounts for highly non-linear, not one-to-one limit surfaces. It allows for the possibility that the same limit might be reached on the same control axis in either direction. Another specific advance is a method for transient limit predictions. This design includes novel prediction and cueing methods for transient limits such as peak flapping. The most significant advance offered with this design is the use of logically based cues. Previous active control system advances are confined to classical control methods used primarily for limit avoidance cues and maneuver envelope enhancement. Only the HACT program is developing more sophisticated guidance cues for station keeping, energy management, and flight path cues 6. This design offers similar guidance and robotic cues using intelligent control systems, specifically fuzzy inference systems for situation identification and fuzzy logic controllers for guidance. The logic-based cues do not rely on crisp mathematical relationships as arithmetically based cues do. These methods complement the tactile cuing and active control concepts advanced in the Helicopter Active Control Technology (HACT) program. These fuzzy reference systems also play a key role dealing with combinations of cues. A final significant aspect of this design is its early development in RIPTIDE. This is one of the first applications of RIPTIDE for novel research and is the first active control program to do so. RIPTIDE provided the results presented here.

3 Cueing Concepts and Terms This design recognizes arithmetically based cues and logically based cues. Arithmetically based cues depend on numerical values. They are calculated from dynamical and probabilistic models and the neural network is the primary tool for these methods. Limit avoidance cues are examples of arithmetically based cues. Logically based cues depend on heuristics. They are inferred from cause and effect relationships, conditional equations, and possibilities. Fuzzy logic is the primary tool for these methods. Procedural cues are examples of logically based cues. Arithmetically Based Cues Arithmetic cues rely on a state space dynamical aircraft model to represent the system of aircraft states, inputs, and outputs. ( ) ( x u) x & = g x, u y = h, ( 1 ) ( 2 ) The state vector, x, is a defining set of aircraft motion characteristics and the input, u, is the vector of physical displacements of the cockpit controls. With information about the states and the controls, and an accurate model of the dynamic interaction between them, the mathematical solution provides the future state of the aircraft. The limited parameters (or limit vector), y, is a vector of individual limits, y i, each of which is an algebraic function of the present states and inputs. Often, a limit parameter is identical to the value of a state. Depending on the context, the word limit may refer either to the name of the limited parameter (such as Vertical Load or Airspeed) or to a critical value of that parameter (such as 4 G s or 150 Knots). The future limit margin is defined as the difference between the limited parameter critical value and the value of that parameter at some future time. y = y y ( 3 ) i fut i lim A control, also called an inceptor, is the physical lever that is the interface between the pilot s applied forces and displacements and the Flight Control System s information medium. The control margin is defined as the difference between the present control position and the critical control position where, if the pilot displaced the controls to that position, the aircraft would reach the critical limit value, the limit. A limit may be a function of the control configuration and flight condition, y lim (x,u), but usually it is a constant maximum or minimum allowable value. A limit has a corresponding upper control margin, when there exists a critical control position greater than the present control position. Likewise, a limit has a lower control margin, when there exists a corresponding critical control position less than the present control position. u = u crit u o ( 4 ) The relationship between the future limit margin and the present control margin is non-causal, non-linear, and non-bijective. To establish a causal relationship and enable practical limit avoidance cueing, every limit prediction model makes a future transition assumption for each limit. With this assumption, the present aircraft state, x o, and the control position, u o, a limit prediction model provides a predicted limit vector, y p (x o,u o ), or predicted limit, y ip. The predicted limit margin is defined as the difference between the predicted limit and the critical limit value or limit. MAX Present } Limit Margin ( y ) i lim ip Map to Control Margin Critical Control Position Control Lever y ( δ ) critical δ o Figure 1. The Key to Effective Tactile Cueing i fut y i p = y i lim y ( 5 ) In a limit avoidance cue, the cueing system approximates a mapping between the predicted limit margin and the present control margin. This mapping of a predicted limit to the critical control position is the essence of effective limit avoidance tactile cueing (Figure 1). i p

4 Logically Based Cues In this design, logical cues rely on fuzzy logic inference systems. The aircraft states, controls, and limits are the fuzzy variables for the inference system. For example, airspeed as a fuzzy variable is not operated on as a numerical value of 60 knots. Instead it is described by membership function such as cruise speed, hover, or below ETL. Likewise, an output such as collective position, has fuzzy membership functions such as forward, centered, or aft. Each membership function is a unimodal possibility distribution across a universe of discourse, analogous to a function domain. A fuzzy inference system follows five steps. First, it fuzzifies the input, converting it from a numerical value into a membership function. Second, it applies the fuzzy operators analogous to the logical AND, OR, and NOT. Third, it applies an implication method. This is a rule described as an IF THEN relationship. Fourth, the results for all the rules are considered simultaneously and aggregated. Finally, aggregate result is defuzzified to number. The rules are defined from expert knowledge such as pilot experience, aircraft technical manuals and handbooks, and aviation textbooks. For example, the rules for an emergency procedure cue are a pilot s answers to: What are the indications that make you realize and identify an emergency condition? and What to you do to remedy the emergency? These become IF-THEN relationships that infer the logical cue. Design Approach and Architecture The Holistic Control System Approach In this design, the active control system is a self-contained and independent control system. It is not a subsystem of the aircraft Flight Control System or even a system strictly in series with the Flight Control System. Rather, it is nested or echeloned within the overall closed loop. (Figure 2). This holistic approach allows pre-existing flight control systems (FCS) and if the active control system is adaptive, it can function properly for different aircraft dynamics and flight control systems. Pilot-Inceptor system in the physical world The Pilot Perception Visual Vestibular Proprioceptive Action Hands On Control Forces Active Inceptor Force Displacement Algorithm Force Counter Force Inceptor Displacement Flight Control System (Digital and Electronic) Flight Control System Control Surfaces Aircraft Dynamics Feedback to the Pilot Feedback to the Active control system Figure 2. Holistic Approach to the Design Feedback to the Flight Control System Aircraft control system designs commonly model the pilot as a subsystem in a closed loop flight control system model. With the addition of an active inceptor, like an active sidestick, the closed system effectively has another feedback loop internal to the pilot subsystem. That feedback loop is the force-counterforce interaction between the pilot and the active sidestick. That feedback channel does not affect the flight control system except by changing the nature of the resulting stick displacement as though the aircraft had a more experienced and knowledgeable pilot. But the active control system still uses input from the aircraft state just as the FCS does. The active control system can also use additional information about the aircraft limits, emergency procedures and other sources that the FCS does not attend to. Tactile cueing can signal the pilot of an impending limit and guide his actions without over-riding them as the FCS would. The pilot retains full authority over the controls.

5 Modular Architecture This active control design is treated as an open engineering system and strives to incorporate robustness through adaptable networks and other methods, and modularity through five functional modules (Figure 3). Each module is set of subsystems with one of five functions: Limit Prediction; Critical Control Position Calculation; Logical Cues; Intelligent Arbitration Among Cues; and Tactile Method Interface. Each of these functional modules is explained in detail. The modular architecture facilitates the use of proven successful systems, such as limit prediction schemes. It also allows additions of new or improved systems and growth into an increasingly comprehensive assistant for the pilot. Limit Prediction Module Each limit parameter has at least one Limit Prediction - Arithmetic Cues Relevant Limit Examples: Vertical Loading Forward & Lateral Airspeed Limit prediction method Math Models Static Neural Networks Type of prediction Dynamic Trim Fixed Time Horizon Torque Angle of Attack Adaptive Neural Networks Flapping Blade Stall Critical Control Calculation Local Sensitivity Methods: Limit-to-Control Partial Derivative Pseudo-Inverse of Limit Gradient Weighted Inverse of Limit Gradient Predicted Limit Surface Search Intelligent Cue Arbitrator Logical Cues Emergency Cues Routine Cues Instrument Cues Pilot Customization Most Conservative Cues Intelligent selection among conflicting cues ( Rule-Based de-confliction and priority ) Relative weighting of controls may vary with limit sensitivity and flight conditions. Tactile Interface Force inversely proportional to control margin Step force at critical control position Shaking (choose magnitude and frequency) Friction Damping Detents Inverse-detents Figure 3. Modular Design Architecture prediction system in this module. Each is defined by three characteristics: the limit, the method of prediction, and the type of prediction. Whether a limit is derived from structural failure criteria, flight control system domain boundaries or regulatory requirements, these arithmetic based systems require a numerically defined limit that depends on aircraft states and controls. Relevant Limits This design loosely classifies aircraft limits into three categories. The category of limits very highly sensitive to control surface movements are appropriately avoided by the aircraft flight control system without cueing to or input from the pilot. A joint U.S./France study 6 of helicopter limits for cueing identified 39 limits that fall into the remaining two categories. These limits are most effectively cued through combinations of visual warning lights and instruments, aural warning and caution tones, verbal (voice) warnings, and tactile cueing through the cockpit controls. The second category, which includes limits that vary slowly over time such as transient limits on the order of seconds, are appropriately cued by non-tactile means. The third category includes the limits that are sensitive to inceptor displacement within a few tenths of a second or are sensitive to inceptor speed within a few hundredths of a second. This module predicts limits for these two categories, primarily the latter. Examples for such limits include: vertical load, main rotor blade stall, main rotor flapping, main rotor speed, and transmission torque. Type of Prediction Two limit prediction methods make two distinctly different types of predictions. The difference is in their assumptions about the aircraft s transition from the present to the future. The fixed time horizon prediction calculates the value of the limit parameter at a fixed distance in the future. In this case, the future transition assumption is an assumed future time history for the controls.

6 The dynamic trim prediction, calculates the limit parameter value for the aircraft dynamical system (equations 1 and 2) in a quasisteady equilibrium. The future transition assumption in this case, is an assumed transition for the states. Fixed Time Horizon (FTH) This type of prediction assumes that the controls will follow some defined path to a chosen point in the future. The fixed time horizon prediction may assume the controls follow the worst-case path. More commonly, the controls are assumed to follow a path similar to the path followed by the pilot during actual or simulated test flights that provide time history data. The fixed time horizon method maps the relationship between the vector of states and controls at each time, t o, to the limit value at time, t o + t. This mapping can be captured in any number of ways, most effectively in neural networks as described below. The advantage of this method is that the time frame for the prediction is known and, depending on the limit, can be reasonably accurate to as far as a half second into the future. x ( t o ) u ( t ) o h y x K u f ( t o + t ) ( t + t ) o K ( t ) K y ( t + t ) K y ( t + 2 t ) K o h f K y = f ( x, u ) p FTH ( 6 ) o Typical time horizons ( t) are 0.25 to 0.46 seconds 7. These prediction times are far enough to give the pilot time to react, but not so far that the prediction loses accuracy. More distant time horizons loose accuracy due to pilot self-determination. That is, the pilot is likely to choose a future control path unlike control path of the aggregate training data for the prediction model. Dynamic Trim (DT) The dynamic trim prediction 8 separates the n aircraft states into k slow states that vary slowly with time, and (n-k) fast states that vary quickly and reach a steady value during a maneuver. x x = x slow fast R n x fast n k R o k x slow R ( 7 ) The future transition assumption is that the controls and the predicted slow states do not change while the fast states have changed and settled to a constant. The predicted limit follows from the solution to the dynamical system (1) and (2) in the form: & slow x 0 = g ( x, u), y = h( x, u) p DT ( 8 ) ( 9 ) The manner in which the fast states transition to steady and the time they take is irrelevant to the method. Consequently, the prediction time is not defined. In practice, the dynamic trim solution can be difficult to find for complex dynamical models, but an adaptive technique 9 can approximate the dynamic trim prediction model from time history a posteriori. The dynamic trim prediction is useful for aircraft in forward flight or in any flight profile where fast and slow states can be discerned. It gives good predictions for the worse case limit values possible during a maneuver. While the prediction time horizon is undefined, this characteristic is generally evident from inspection of the time history of the prediction and the actual limit value. Limit Prediction Methods Math Model This prediction method uses a model for predicted limit, y p, derived from a priori understanding of the aircraft dynamics. y p = f ( x, u) ( 10 ) This method solves the state equation (1) based on the future transition assumption. For the dynamic trim prediction, the assumption defines values for the fast states. For the fixed time horizon prediction, the assumption defines the control future time

7 history. The one special form of the math model that requires no future transition assumption is the zero time horizon prediction, which is not a prediction at all. In that case, the present limit is used as the prediction, y p =y. The math model produces a virtual table of limit predictions for given states and control values. This can take the form of an actual look-up table for use with multiple argument interpolation, but more commonly this prediction method is a preliminary step to create neural network training data. Static Neural Networks An artificial neural network is a mathematical construct, such as a polynomial or a combination of vector functions called basis functions (such as the sigmoid, tan-sigmoid, and radial basis functions). Based on error back-propagation, this construct has parameters that self-adjust to provide a target output. Neural networks capture the a posteriori relationship between the controls and the predicted limits based on representative pattern and target data. Math model solution sets can provide this data directly or the time history data from flights and simulations can provide it. Static network training is completed with all the data available at once. Type Training Patterns Training Targets Dynamic Trim x slow (t), u(t) f NN (x,u) y DT (t) Fixed Time Horizon x(t), u(t) f NN (x,u) y(t+ t) p ( t) f ( x( t) u( t) ) y =, ( 11 ) NN Prediction error is a common practical difficulty with math model and static neural network predictions because aircraft parameters and flight conditions change, as when the center of gravity shifts or pilot control characteristics change. The HELMEE and HACT projects correct prediction errors using complementary filters that effectively eliminate steady state prediction errors. But while this technique performs an essential function, the filters cloud the output from the prediction model. Adaptive Neural Networks Adaptive neural networks offer an alternative method to correct real time prediction errors and, unlike filters, they improve the prediction function to capture local or transient variations in the dynamical relationship of states, limits, and controls. Unlike a static network, an adaptive network adjusts the neural network weights incrementally, as additional pattern and target pairs are presented. In other words, the adaptive neural network uses time history data in real time to reduce the prediction error and improve the prediction model. In order to use an adaptive network to approximate the predicted limit, it must have a measured or inferred value for the limit parameter to use as its real time target. Alternatively, the adaptive net could model local state dynamics as shown in Figure 4, instead of the limit parameter. In this method, the limit parameter is modeled as a dynamic system using a mathematical model or a static neural network trained with representative time history data. The Approximate Model captures the dynamic complexity (high order characteristics) of the limit parameter. The Approximate Model should be a good global representation of the parameter dynamics. Given a powerful enough control computer, the Approximate Model could be a Helicopter Dynamics x u x ˆ 0 = x u* Approximation Math Model xˆ & = g( x, u) or Static Neural Net ( ) x & ˆ = f x, u NN xˆ N + 1 Forward Euler Method x &ˆ Approx. ( ) + x & ˆ x & ˆ Model = t + Adaptive Net ( ) y = h p x ˆ, u K Adaptive Net f y ANN lim ( x, u) ˆx N + - e x u ŷ Figure 4. Adaptive Neural Net Application with a Transient Limit

8 comprehensive math model such as GenHel. The Adaptive Neural Net (ANN) supplements the approximate model with a reduced order correction to correct for local and short-term variations in the parameter dynamics. The ANN is kept simple for two reasons: 1) to reduce computational demands for real time adaptation and 2) to prevent the ANN from dominating global dynamics with local and short term dynamics. For fast transient limits, this dynamical model must be accurate for short time predictions, on the order of 0.05 seconds, but the model need not be accurate beyond that time horizon. With a dynamical model for the transient limit parameter (made from the combination of the approximate model and the adaptive neural net) this prediction scheme uses a forward Euler approximation to model the future response of the parameter. The system tracks the maximum and minimum values for the modeled parameter to find the peak. Here, the future transition assumption exists in the choice of the future time history of the control, u *. A reasonable assumption for u * could be the fastest and farthest the pilot could realistically move the control within the time horizon. This prediction algorithm could be repeated with several variations for u * and the worst likely result chosen for the prediction. The time frame for fast transient limits is too brief for effective pilot cueing with a force that depends on control position (a softstop ). Transient limits like main rotor flapping with respect to cyclic and vertical loading with respect to collective are more strongly influenced by the speed of control movement rather than control position. A more effective cueing method alters the inceptor damping. The limit margin, y, is defined as the difference between the maximum allowable transient peak and the predicted peak for the limit parameter. The damping cue is proportional to the worst-case fast transient limit margin. Local Sensitivity Methods Critical Control Position Module Each system in this functional module establishes a relationship between a limit and the controls. Local sensitivity methods depend on the limit gradient or the limit vector Jacobian, also called the limit sensitivity matrix. This method approximates a linear limit-to-control relationship using the tangent to the limit surface defined by the math model, y p =f(x,u), or neural network y p =f NN (x,u). If the limit prediction function is well understood, the predicted limit Jacobian can be found analytically. If not, the local limit sensitivity is found through perturbation methods, iterating on its limit prediction system as a subsystem or subroutine. For the non-predictive limit models, y p =y, the critical control position equals the current control position, u crit =u o These methods have the advantage of computational speed. The disadvantages are those inherent in the linearization. The limit surface may be highly nonlinear and local sensitivity values may vary considerably with small changes in the state or control. Also, it is not uncommon for the current control position on the limit surface to lie at a local minimum or maximum where the same limit is reached whether a control is moved one direction or the opposite. Linearization will fail to predict accurate critical control positions for these conditions. Limit-to-Control Partial Derivative This simple method finds the control margin by dividing the limit margin by the limit sensitivity for each control axis (Figure 5). u j f i = u j 1 ( y y ) i p i ( 12 ) This limit sensitivity method estimates the critical position for each active axis independently. The HELMEE study used this method effectively to cue each limit along a distinct active control axis. lim Limit Parameter vs. Two Active Control Axes 100% Control Position u 1 Current Control Position u 1 Local Limit Surface u 2-100% Critical Position -100% Control Position u 2 100% Figure 5. Critical Position from Limit Partial Derivative

9 Pseudo-Inverse of Limit Gradient An alternative method 8 treats the controls together as a vector and uses the Jacobian s pseudo-inverse to find the control margin vector to the nearest control combination that zeros the limit margin. This nearest distance is the least-squared distance of each axis control margin. This method weights each of all the controls equally. fi u = + ( y y ) Limit Parameter vs. Two Active Control Axes 100% u lim ( 13 ) i p i u -100% Critical Position The critical control position for each axis follows directly from Control Position u 2 Figure 6. Pseudo-Inverse of Limit Gradient. the control margin vector decomposition. This method works fairly well when one limit is moderately influenced by two or more active control axes. Weighted-Inverse of Limit Gradient Control Position u 1 Current Control Position Local Limit Surface A variation of the previous method multiplies the pseudo-inverse by a weight matrix. This weight vector may be a function of the states to emphasize or de-emphasize control axes at different flight conditions u = W x Predicted Limit Surface Search + fi ( ) ( y y ) u i p i A new method developed for this active control system design uses a limit surface search algorithm to find the critical control position. This method begins a search at the current control position, x o, and samples the prediction models, y p (x o,u), at increasing and decreasing positions for each of the active control axes in turn. Throughout the search the present (instantaneous) state vector is used. When the resulting lim n x R k u R W : R prediction for a limit first moves into its set of prohibited values, that control position becomes the critical control position. A prohibited value for the limit parameter is one beyond the maximum or minimum allowable or may be a dangerous or damaging internal subset of values. Current Control Position This method does not assume a positive or negative Current Nz ( δcoll, δlong, y ) relationship between the control and the limit. It does allow Predicted Nz value ( δcoll, δlong, yp ) Figure 7. Predicted Limit Surface Search Algorithm. the possibility that the non-linear inverse may not be one-toone. It has these advantages over the local sensitivity methods described earlier. Its chief drawback is its computational demand. Without a capable active control system computer, the designers may need to simplify the complexity of the neural network used for the limit prediction or reduce the resolution of the limit surface search. The latter option is usually best since the prediction is itself only an approximation and there is no need to search to high precision a limit surface of lower precision. n R The Mesh Surface of Figure 7 represents a predicted limit (vertical loading, N z ) with respect to collective and longitudinal control axes. At the depicted instant in time, during a pull up maneuver, when the control and limit coordinate is positioned at (δ coll, δ long,y), the search algorithm begins at the predicted limit for the current control position (δ coll,δ long,y p ). The algorithm k k ( 14 )

10 varies each control position in the prediction function away from the start position, along the admissible control positions shown as black lines. When the prediction exceeds the limit (in this example y lim+ = N z(max) =1.5), that control position is defined as the critical control positions for each axes for that instant. Those critical upper critical control positions are indicated in red and blue lines. Note in this example that the predicted limit decreases at very high collective positions. Had the limit been set a little higher (i.e. N z(max) =1.6), the algorithm would not find a critical position for collective because no position along the collective search path exceeds the limit. Logical Cues Module These logically based cues do not rely on numerically defined limits and arithmetically determined cues. Instead, fuzzy logic inference systems serve as expert controllers and the active control system effectively becomes a robot assistant. These cues assist the pilot with routine tasks, prompt him to initiate emergency procedures, assist with instrument flight, and customize the active controls to individual pilots. In many cases, these cues are translated to force characteristics qualitatively different from limit avoidance cues and do not interfere with them. In other cases, the arbitrator must prioritize them among limit avoidance cues. Emergency Procedure Prompts Emergency prompt systems integrate two functions: emergency situation identification and tactile prompting for immediate action. Fuzzy logic heuristics detect, identify, and verify the emergency condition using indications delineated by the aircraft operator s manual emergency procedures section and other expert sources. These heuristics also consider the pilot s action or inaction and assess the criticality and appropriateness for an emergency prompt. When an emergency condition is verified and critical timely action is required, the second function makes an appropriate force prompt for the pilot. Vortex Ring State Avoidance If the vortex ring state could be well defined as a numerical limit, an arithmetically based limit avoidance cueing system would apply. The HACT Program takes that approach to provide a power settling avoidance cue on the collective control axis. However, when the condition is not explicitly defined but is generally understood, an expert model assesses the possibility of the condition and sets tactile avoidance cues and non-tactile cues. This vortex ring avoidance cue treats the condition not as an arithmetic cue as does the HACT program, but as a logic based cue. While not usually addressed in helicopter operator s manuals, flight schools include settling with power as an important topic of instruction. School manuals 10 describe the conditions conducive to settling with power as: a vertical or nearly vertical descent of at least 300 feet per minute, low forward airspeed, and normal-high engine power (from 20 to 100 percent). From this knowledge (depicted in Figure 8), an abridged fuzzy inference system takes the form: Rate of Descent Member Functions Low Steep Fast Autorotative Boundary Vortex Ring State Descent Angles Low ETL Cruise Vmax Horizontal Speed Member Functions Figure 8. Fuzzy Inference of Vortex Ring State. Operator Fuzzy Variable Membership Function IF Torque is Nominal (20-100%) AND Rate of Descent is Steep ( fpm ) AND Horizontal Speed is Low ( <25 kts) THEN Vortex Ring is Imminent With the condition identified, the fuzzy cueing system simultaneously acts as a fuzzy controller for an active control axis cue. Again, flight school instruction can form the basis of the rules for the collective and cyclic cues shown below:

11 The abridged fuzzy logic controller for an avoidance cue when vortex ring state is imminent: Operator Fuzzy Variable Membership Function WHILE Vortex Ring is Imminent AND Altitude is Safe THEN Collective Cue is Decreasing Operator Fuzzy Variable Membership Function WHILE Vortex Ring is Imminent AND Altitude is Safe THEN Longitudinal Cue is Decreasing (Forward) AND Lateral Cue is Increasing (left or right) Autorotational Entry Prompt Many high performance rotorcraft use high disk loads and low blade inertia to enhance agility. Such rotor systems loose rotational speed very quickly when power fails due to engine failure, drive shaft failure, or similar emergencies. In these situations, the pilot must immediately identify the emergency, reduce collective to maintain rotor speed and enter an autorotational descent. An autorotational entry prompt assists the pilot by identifying the emergency and providing a downward cue on the collective. If the pilot had set a friction to the collective, the active control system would disable the friction. If the pilot had removed his hand from the collective, the force prompt would physically lower the collective far enough to maintain safe rotor speed. An autorotation entry cueing system would only prompt the pilot to take immediate action. It provides no prompt when its heuristics deem such action inappropriate. This might occur at Operator Fuzzy Variable Membership Function low altitudes (below 100 feet) or when the pilot is already taking IF Ω Rotor is Below Nominal corrective action. Any emergency procedure cue must include OR d/dt ( Ω Rotor /Ω Turbine ) is NOT Zero thoughtful rules to prohibit interference with the pilot during emergency maneuvers. Consequently, in an emergency situation, AND Ω & Rotor is Decreasing the cue arbitration module may disable all limit avoidance cues. AND Altitude is NOT Low THEN Collective Cue is Decreasing Routine Cues AND Friction is Zero Position Cue for Hover An active cue hover hold takes the form of a force detent or a light preload force, described in more detail with the tactile interface module, to physically manipulate the active control inceptor and keep the aircraft in a fixed position. As with other tactile control cues, the pilot can immediately override it by applying a force greater than the cueing force. An active hover cue can operate without interfering with the pilot as he maneuvers from one position to another. But whenever the aircraft approaches a hover, the pilot feels the cyclic fall into the force detent. He could then release the cyclic and allow the hover cue to hold the aircraft stationary at a hover. The rule base for a fuzzy logic controller depends primarily on position, speed and angular rates. Decoupled Axis Cueing When the position cue rule set decouples the longitudinal, lateral, and vertical cues to respectively drive the longitudinal cyclic, lateral cyclic, and collective; the fuzzy logic controller provides an axis position cue. An energy management cue is an example of such a cue. It guides the pilot when decelerating to and accelerating from a hover without altitude loss or gain. In this case, the pilot follows the collective cue (or lets the cue physically move the collective), while he moves the cyclic away from its hover cue. The vertical cue will physically adjust the collective to maintain a constant height above ground while the aircraft moves horizontally. Another form of this cue is the vertical mask/unmask cue. Here, the pilot follows the cyclic cues (or allows the cue to physically position the cyclic), while he increases and decreases collective to unmask and mask the aircraft. The active control system physically moves the cyclic to maintain the same horizontal position. The final example is the lateral mask/unmask cue where the pilot allows the active control system to move the collective and longitudinal cyclic while he changes the lateral position.

12 Instrument Cues Spatial Orientation Cueing During instrument flight, pilots force themselves to ignore outside visual information and proprioceptive perception of the aircraft s spatial attitude. Trust Your Instruments is the mantra and pilots maneuver the aircraft solely on the basis of the visual information provided by the artificial horizon, altimeter, turn and slip, and other instruments. Instrument flight sometimes demands great force of will to ignore the gut feeling that the aircraft is flying awry. Pilots must overcome the combined power of vestibular and proprioceptive misperception. An active control system can augment the instrument panel indications with supporting tactile cues. The pilot would then have mutually supporting visual and proprioceptive indications and can more easily overcome his conflicting vestibular and remaining proprioceptive cues to recognize the true spatial orientation of the aircraft A spatial orientation cue uses inceptor displacement and force to respectively command rate and attitude. In such a scheme, the aircraft flight control system accepts physical displacement of the longitudinal and lateral control inceptors as its input for pitch and roll rate commands. The active control system provides a counter-force as a function of the aircraft s pitch and bank angles. So a constant right lateral cyclic displacement would command a constant roll rate, but a pilot s constant lateral force would command a constant bank angle. Or alternatively, if the pilot provided no lateral force or took his hands off the control stick, the active control counterforce would command the displacement and bring the aircraft to a wings level attitude. Consider the example of an instrument pilot making a right turn. When he initiates the turn from straight and level flight, the lateral stick is centered and neutral. The pilot moves the stick right and feels the nominal leftward counterforce just as he would with a passive inceptor. The flight control system recognizes the displaced inceptor and commands the control surfaces to roll the aircraft right. As the aircraft begins to bank right, the pilot feels the counterforce increase. This force tells the pilot that he is in a right bank and gives him the intuitive feeling that the aircraft wants to bring itself back level. If the pilot held the rightward displacement, the aircraft would continue to roll right into a barrel roll and the left counterforce would increase to a maximum when the aircraft reached a 90 right bank. If the pilot allowed the force to guide him, he would hold a constant right force against the stick, but the stick would move left to a centered position where the flight control system would recognize zero displacement and command the aircraft to a zero roll rate and the bank angle would be constant. When the pilot decides to terminate the turn and return to wings level, he would reduce his force against the stick to zero or simply remove his hand altogether. The active control system force would physically guide the stick to the left until it reached a zero force position with a left displacement. The flight control system would recognize the left displacement and command a left roll rate. As the aircraft rolled left towards a wings level attitude, the active control system recognizes the decreasing bank angle and moves the zero force position of the lateral stick to the right. This zero force position guides the stick to the right until it reaches the centered position when the bank angle is zero. Instrument Approach Cueing Similar active cues can guide the pilot during an instrument approach. An instrument approach active cue moves the zeroforce positions of the active control axes to command the aircraft along the proper glide slope and approach course. The active control system may have its own approach autopilot system, but some advanced flight control systems already provide an instrument approach autopilot. Such a flight control system could provide the control displacement algorithm to the active control system instead of back driving the control position. The pilot can follow the tactile cue while observing the conditions outside the cockpit. When the aircraft descends below the cloud base, the pilot can immediately transition to visual flight and complete the approach, which may include timely and delicate maneuvers for low decision heights, circle to land approaches, and other challenging terminations. A tactile approach cue can ease the sometimes disorienting transition from instrument flight to visual

13 flight. With effective spatial orientation and instrument approach tactile cues, the pilot could conceivably conduct a no-gyro radar precision approach blind folded. Pilot Customization Passive Customization An active inceptor can define the nature of the nominal force displacement relationship to suit pilot preferences and needs. Even passive inceptors allow pilot adjustments though cockpit knobs and buttons for collective and cyclic friction, fore to aft cyclic centering, and force trim. An active control system allows similar presets through instrument panel interface menus or directly through switches beside the inceptor. The range and variety of force-feel characteristics vary with the design of the active controller but commonly include static friction, dynamic friction, stiffness, damping, and range of motion. Passive customization allows the pilot to select these tactile qualities to suit his preferences. Active Customization and Pilot Induced Oscillation Active customization automatically adjusts global force characteristics, primarily damping and stiffness, to ameliorate pilot aggressiveness, over-controlling, and pilot induced oscillation. Pilot-aircraft dynamic interaction is the basis for Pilot Induced Oscillation (PIO). An active customization system identifies patterns of unfamiliar or overly aggressive pilot behavior and adjusts the tactile quality of the cockpit controls. Such flying patterns range from benign over-controlling and porpoising to dangerous limit-cycle oscillations involving the pilot, the aircraft dynamics, and the control surface actuators. Rigorous design standards 11 for aircraft and flight control systems eliminate the latter, dangerous form of PIO. An active control system can potentially eliminate or ameliorate the remaining forms of PIO. A fuzzy logic inference system can identify pilot behaviors indicative of PIO, such as a sudden increase in power in the time history of control and state frequency content 12. Once identified, the system can positively affect the time-constant of the physical interaction between the pilot and the inceptor. The natural frequency and damping define the apparent inertia of the inceptor and can make the aircraft feel more or less responsive. It gives the pilot a sense of how rapidly the aircraft can react to his movements. This is analogous to the different experiences felt from wagging a pencil in hand compared to the feeling of wagging a brick. Cue Arbitration Module With multiple critical control position vectors, u crit, for multiple limits and logical cue positions across multiple control axes, the design needs an arbitrator to decide which cue among many will drive each style of tactile cue for each of the active control axes. This module defines the cue position, u cue, which the tactile interface module will use. In most instances, the solution is the most conservative cue for each control axis. But it is not always appropriate to cue every control for the most conservative limit. At times the cues may conflict with one another as when one limit is exceeded because a control axis is too far left while another limit is exceeded because the same control is too far right. In such cases, the arbitrator uses a rule-based method of de-confliction and appropriate cue selection. Depending on the precision or confidence of the control cue, the arbitration module may command an abrupt step force cue for high confidence limit predictions or a more gradual cue for less well-defined or low confidence predictions. Most Conservative of Several Limit Cues Arbitrating among multiple cues may be very simple. Usually the most conservative control position of the multiple limits should drive the tactile cue. For example, consider a moment of forward flight when the longitudinal cyclic position is forward at -

14 5%. The Critical Control Positions for two limits are 30% aft for vertical load limit and 45% aft for the main rotor blade stall limit. The most conservative method chooses 30% aft as the combined critical control position. Control Axis Selection When each limit is invariably mapped to a primary control axis, the most conservative method is straightforward and simple. But a limit s relationship to the control axes changes with the flight condition, a fuzzy inference system decides which control axes are most appropriate as cues for each limit and whether the cue should be tactile or non-tactile. The system eliminates the cue or limits it to a subset of permissible positions. The primary indicator of an appropriate control axis is the dynamic nature of the critical control position for the axis. When the critical position changes rapidly across a large range of control positions, it is less desirable as a tactile cue. This generally occurs when the limit sensitivity is low with respect to the control or when the limit varies rapidly due to the states. In such a case, the limit is too fast to cue. A force cue seems jittery and unpredictable and the pilot is likely to find it objectionable. In this case, the limit arbitrator entirely eliminates the cue from the control axis. A second form of selection restricts the cue to a subset of the control positions. The rules that shape this inference system rely on the knowledge interrelated limits and the consequences of control positions. These rules are specific to the relevant limits of the prediction module. Examples of interrelated limits are vertical loading and main rotor blade stall. When the aircraft approaches those limits together, as in a pull up maneuver, both avoidance cues would push the longitudinal collective forward. In extreme cases, the cue would push the collective forward and put the aircraft into a dive that would exacerbate the problem. Intelligent Selection of Conflicting Cues The cue arbitration module uses a continuum of fuzzy logic assigned weights to emphasize or ignore each active control axis. This effectively prioritizes the urgency of the limits. In cases when the aircraft flies beyond two or more limits simultaneously, the critical control positions may be conflict. The critical control position for one limit may be above the current position while the critical position for different limit may be below the current position. The arbitrator emphasizes the cue for the most urgent rule. Tactile Interface Module After the cue arbitrator decides the one set of cue positions, the tactile interface converts the information into intuitive force cues for the pilot. In general, the force cue is a function of the cue positions from the cue arbitration module, the position of the controls, the velocity of the controls, and the time. The cueing force is a combination of the nominal force displacement curve, softstops, the detents, oscillations, damping and natural frequency response. Because human pilots have different degrees of strength and control for the different control axis, it is appropriate to decompose this function into its active control axis components and tailor them to pilot physiology. ( u,u,u,t) = F + F + Fdet + Fω + Fω ζ Ffric = F cue nom ssi i j & ( 15 ) F n + Nominal Force-Displacement Relationship (F nom ) An inceptor uses a nominal force-displacement relationship where the pilot feels a centering force that increases gradually and nearly linearly as it is pushed away from its neutral position. nom ( u) = mu Fo F F + j = ( 16 ) The zero-force intercept is the neutral position where the inceptor will settle when left untouched. An active sidestick can offer cues and guidance by changing the zero-force position and how the counter-force increases as pilot applies force. The forcedisplacement relationship can be nearly flat, m=0. This is the typical feel of a traditional helicopter cyclic stick without friction.

15 Another relationship uses a preload force. With a preload, the control will not move from the neutral position until the breakout force is reached. Another significant choice is the inceptor range of displacement. Traditional cyclic sticks move several inches in two axes. An active sidestick may move approximately 25 degrees or more longitudinally and laterally and may provide a third axis in the twist about a vertical axis. Smaller ranges of movement, such as 5 or 10 degrees from neutral, are useful when force is the only interaction between the pilot and the active system. Larger ranges of movement, such as 15 to 30 degrees from neutral, allow both force and displacement as information channels between the pilot and the active control system. However, very large ranges of movement in a sidestick can be awkward for the pilot. Also, a larger range magnifies the movement of limit avoidance cues to the point where they may be objectionable to the pilot. The range of movement may best be left adjustable for pilot preference. Force Inversely Proportional to Control Margin (F ss ) One tactile cueing method is the use of a Force Inversely Proportional to Control Margin (Figure 9). This form of a softstop, used successfully with V-22 simulations, creates a counter force that opposes the pilot as he pushes the control towards a limit. The magnitude of the counterforce is approximately inversely proportional to the control margin and increases to a maximum counter-force at the critical control position. This method can be implemented with minor variations, but its defining characteristic is the gradual increase in counter-force as the critical control position is approached. This method does not provide a decisive cue regarding the limit and this reflects the true indistinct nature of many (perhaps most) limits, which are based on subjectively defined safety margins added to structural failure loads or control system domain boundaries. Step Force at Critical Control Position (F ss ) Another successful form of softstop uses a step increase in counter-force at the critical control position (Figure 10). This is the primary cueing method for the RASCAL active control system because it provides a decisive indication to the pilot about the location of the edge of the flight envelope defined by the limit prediction algorithms. However, when the critical control position varies rapidly while the pilot is following the cue, it can seem jittery and may be objectionable. Detents and Inverse-Detents (F det ) A force detent superimposed on the nominal force-displacement relationship serves well as a trim cue or an autopilot cue. The sidestick will remain in a detent force-well until the pilot provides a sufficient break away force (Figure 9). Then the stick would follow the nominal force-displacement relationship. The inverse detent has the opposite effect (Figure 10). It pushes the stick away from the inverse-detent position to one side or the other. Such a cue steers the pilot away from high-risk flight conditions, such as very steep, high power approaches where vortex-ring state is predicted as imminent. Applied Force Steepness Force Displacement Relationship How it feels Force Detent Control (inceptor) Travel Force Detent for guidance Force F 1 u } Force Cue Figure 9. Force Inverse to Control Margin with Detent. Steepness Applied Force Force Displacement Relationship How it feels Inverse Force Detent Inverse Detent for avoidance Step Force At u crit Control (inceptor) Travel Force Cue Figure 10. Step Force at Critical Position with an inverse detent.

16 Shaking and Vibration (F ω ) Shaking and vibration is a very useful supplemental cue. It is used to indicate that the aircraft is already beyond a limit. It can also cue impending limits whose indications involve vibration. For example, a high frequency vibration can cue loss of tail rotor effectiveness and tail rotor malfunctions. A low frequency, 1/rev, can cue main rotor stall and other main rotor limits. ( t) = Asin( ωt) F = F ( 17 ) Damping and Natural Frequency Response (F ωnζ ) The frequency response of an active inceptor can imply agility or sluggishness to convey the maneuvering capability of an aircraft in varying flight regimes. Damping as a force cue, can be very effective for transient limits such as maximum flapping with respect to cyclic. It is the only force cue listed here that depends directly on control speed. Maximum transient limits depend primarily on fast control movements rather than control positions. Friction (F fric ) Fω ς = M ( u&& + 2ωnς u& + ωn n Friction is a constant force that opposes the direction of movement. It may have use as a cue, but mainly it helps the pilot hold the control at a constant position despite airframe vibrations or those occasions when the pilot removes his hand. 2 u ) ( 18 ) Limit Avoidance Cueing for the RASCAL Helicopter Applications Prototype development for the RASCAL active control system with the Sterling Dynamics Active Sidestick System model SA-S-2D-1 began in the summer of 2001 at the Army/NASA Rotorcraft Division within the Real-Time Interactive Prototype Technology Integration/Development Environment (RIPTIDE), which the Rotorcraft Division designed for just this sort of project 13. The RIPTIDE uses SIMULINK as a control system development tool and provides a simulation environment using any of several math models, including the UH-60A general helicopter (GenHel) model. The first application emulated the previous success of the HELMEE project. The active control system had the structure depicted in Figure 11. No logical cues were used. The Main Rotor Stall limit was defined numerically as Equivalent Retreating Indicated Tip Speed (ERITS). ρ ΩR V ρo ERITS = N z eq W W o ( 19 ) The prediction model was the same polynomial static neural network developed for the HELMEE study. It provided a prediction for a fixed time horizon of seconds. A complementary filter between the neural network and the instantaneous ERITS value eliminated steady state prediction error. ERITS values below 250 were considered beyond the limit, and were signaled by a shaker cue. An ERITS prediction of 300 defined the placement of the softstop. y = 300 fps lim ( ERITS) ( 20 ) Limit Prediction Critical Control Calculation Intelligent Cue Arbitrator Tactile Interface Relevant Limit Main Rotor Blade Stall Limit prediction method Static Neural Network with Complementary Filter correction Type of prediction Fixed Time Horizon Limit-to-Control Partial Derivative Most Conservative Cue (direct single axis cue) Step force at critical position Shaking (beyond limit) Figure 11. Main Rotor Blade Stall Cueing

17 (Aft) 25 Pilot overrides blade stall limit cue Pilot follows the blade stall limit cue Longitudinal stick (deg) 0 (Fwd) -25 Stick Position, δ Critical Position, δ crit Stick Position, δ Critical Position, δ crit 800 Actual ERITS, y Actual ERITS, y ERITS (fps) Predicted ERITS, y p Predicted ERITS, y p y lim = 300 y lim = Time (seconds) Time (seconds) Figure 12. Main Rotor Blade Stall limit avoidance cueing in piloted RIPTIDE simulation. The performance of the active control system in piloted simulation * of two consecutive pull-up maneuvers is shown in Figure 12. The position of the softstop and stick are shown in the top graphs. The predicted control margin is the area below the softstop (in red) and above the stick position (in black). In both maneuvers, the aircraft begins in an accelerating dive where the limit parameter, ERITS, is approaching its limit. Consequently, the control margin is narrowing. When the predicted ERITS reaches its limit as the stick moves aft, the pilot encounters the softstop cue. In the first maneuver (at left), the pilot overrides the cue to make an abrupt pitch up. He exceeds the limit as ERITS drops to 175 fps. At critical times, the pilot may need to do this to avoid sudden obstacles (i.e. wires) and an active cue does not prevent him. In the second maneuver (at right), the pilot encounters and follows the cue, and in so doing gets the most out of the maneuver envelop without significantly exceeding the limit. * This piloted simulation is available as a QuickTime movie at the author s website,