Rotor torque predictor

Abstract

A system is disclosed for determining the total torque required at the main and tail rotors of a helicopter for use in feed-forward rotor torque anticipation which includes a polynomial neural network adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon a plurality of pilot inputs and airframe inputs, a main rotor load map for determining the torque at the main rotor in hover out of ground effects based upon main rotor speed and collective stick position, a tail rotor load map for determining the torque at the tail rotor with the helicopter stationary based upon main rotor speed and pedal position, and a processor for calculating the required total torque at the main and tail rotors by summing the outputs of the polynomial neural network, and the main and tail rotor load maps.

Description

BACKGROUND OF THE INVENTION

[0003] 1. Field of the Invention

[0004] The subject disclosure relates to control systems for helicopters, and more particularly, to a system for predicting rotor torque demand to aid in rotor speed control.

[0005] 2. Background of the Related Art

[0006] To improve engine power management and aid in rotor speed control, modern helicopter engine control systems such as Full Authority Digital Engine Control (FADEC) systems utilize complex algorithms to anticipate a change in power demand. These anticipation algorithms utilize a number of pilot and airframe inputs including collective stick rate, yaw control rate, lateral stick rate and rotor speed rate to predict torque requirements. This feed-forward anticipation approach controls fuel flow to accelerate or decelerate the engine, thereby preventing rotor droop or an over-speed condition from occurring as a result of sudden torque inputs imposed on the rotor blades. There is a need however, to simplify the algorithmic approach presently used to anticipate rotor demand on the engine and reduce the number of inputs required to predict the demand.

SUMMARY OF THE INVENTION

[0007] The subject invention is directed to a new and useful system for determining the total (aerodynamic+static) torque required at the main and tail rotors of a helicopter for use in feed-forward rotor torque anticipation to aid in rotor speed control. The system includes means for predicting the aerodynamic torque at the main and tail rotors with the helicopter in motion, means for determining the torque at the main rotor with the helicopter stationary, means for determining the torque at the tail rotor with the helicopter stationary, and means for calculating the required total torque at the main and tail rotors based upon the aerodynamic torque at the main and tail rotors with the helicopter in motion, the torque at the main rotor with the helicopter stationary, and the torque at the tail rotor with the helicopter stationary. Preferably, the means for calculating the required total torque at the main and tail rotors includes means for summing the aerodynamic torque at the main and tail rotors with the helicopter in motion, the torque at the main rotor with the helicopter stationary, and the torque at the tail rotor with the helicopter stationary.

[0008] Preferably, the means for predicting the aerodynamic torque at the main and tail rotors with the helicopter in motion comprises a polynomial neural network. In accordance with the subject invention, the polynomial neural network may be a full order polynomial neural network or a reduced order polynomial neural network depending upon the number of candidate inputs. Alternatively, the system may be configured with means for selecting between a full order polynomial neural network and a reduced order polynomial neural network.

[0009] The full order polynomial neural network is adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon pilot input data including collective stick position data, pedal position data, rate of change of pedal position data, lateral stick position data and longitudinal stick position data, and airframe input data including main rotor speed data, engine torque data, true airspeed data and pitch attitude data. In contrast, the reduced order polynomial neural network is adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon pilot input data including collective stick position data and pedal position data, and airframe input data including main rotor set speed data and climb rate data.

[0010] Preferably, the means for determining the torque at the main rotor blades with the helicopter stationary is a main rotor load map, and main rotor load map is adapted and configured to determine the torque at the main rotor with the helicopter stationary based upon main rotor speed and collective stick position. Similarly, the means for determining the torque at the tail rotor blades with the helicopter stationary is a tail rotor load map, and the tail rotor load map is adapted and configured to determine the torque at the tail rotor with the helicopter stationary based upon main rotor speed and pedal position.

[0011] In sum, a system is provided for determining the total torque required at the main and tail rotors of a helicopter that includes a polynomial neural network adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon a plurality of pilot inputs and airframe inputs, a main rotor load map for determining the torque at the main rotor in hover out of ground effects based upon main rotor speed and collective stick position, and a tail rotor load map for determining the torque at the tail rotor with the helicopter stationary based upon main rotor speed and pedal position, and means for calculating the required total torque at the main and tail rotors by summing the outputs of the polynomial neural network, and the main and tail rotor load maps.

[0012] These and other aspects of the system of the subject invention will become more readily apparent to those having ordinary skill in the art from the following detailed description of the invention taken in conjunction with the drawings described herein below.

BRIEF DESCRIPTION OF THE DRAWINGS

[0013] So that those having ordinary skill in the art to which the subject invention appertains will more readily understand how to make and use the same, reference may be had to the drawings wherein:

[0014] FIG. 1 is a functional block diagram of a rotor torque anticipator configured in accordance with a preferred embodiment of the subject invention which is associated with the aircraft flight control computer and the engine fuel control system;

[0015] FIG. 2 is a schematic representation of a reduced order polynomial network configured in accordance with a preferred embodiment of the subject invention having ten (10) interconnected nodes each having a quadratic polynomial associated therewith;

[0016] FIG. 3 is a graphical representation of comparative rotor torque predictions resulting from a simulated pull-up from a powered decent;

[0017] FIG. 4 is a graphical representation of comparative rotor torque predictions resulting from a simulated autorotation recovery to level flight; and

[0018] FIG. 5 is a graphical representation of comparative rotor torque predictions resulting from a simulated high speed terrain avoidance maneuver.

DETAILED DESCRIPTION OF THE PREFFERED EMBODIMENTS

[0019] Referring now to the drawings wherein like reference numerals identify similar aspects of the system disclosed herein, there is illustrated in FIG. 1 a schematic representation of a feed-forward rotor torque anticipator configured in accordance with a preferred embodiment of the subject invention and designated generally by reference numeral 10. In general, rotor toque anticipator 10 is adapted and configured to predict the demanded rate of change of the gas generator speed based upon the total (aerodynamic+static) load torque at the rotor blades, so as to minimize transient rotor speed droop and/or overshoot as a result of sudden torque inputs imposed at the rotor blade by abrupt aircraft maneuvers that demand large and/or rapid power excursions. This anticipation logic is configured to bypass the rotor/power turbine speed governor and accelerates/decelerates the engine as required without waiting for rotor/power turbine speed errors to build-up.

[0020] In accordance with the subject invention, the total load torque at the main and tail rotor blades of the helicopter is determined by the aircraft flight control system and is composed of three parts. These include: 1) the torque at the main rotor blades with the aircraft stationary, QROTmain; (2) the torque at the tail rotor blades with the aircraft stationary, QROTtail; and (3) the additional (transient) torque at the main and tail rotor blades with the aircraft in motion, QROTaero. These three components are the most instantaneous estimates of load torque, because they exist at the rotor blades. Torque measurements in the shaft of the rotor/drive train between the engine and the rotors is delayed in time by the inertia of the rotor blades. In other words, engine output shaft torque transiently lags the torque at the main and tail rotor blades, and therefore is not suitable for use as a load anticipator. The total instantaneous torque at the rotor blades, QROT, is therefore the best signal to be used as a load anticipator, because it provides the greatest lead to the engine control system

[0021] The static main rotor torque and the static tail rotor torque are well known because they can be readily measured by the engine shaft torque sensor when the aircraft in hover. Therefore, these measurements can be directly programmed into the flight control computer of the aircraft as a function of rotor speed and blade pitch. Accordingly, the flight control computer stores data representing two engine performance load maps for generating torque values based upon specific pilot and airframe input signals. One load map is the Main Rotor Load Map in Hover 12 and the other load map is Tail Rotor Load Map in Hover 14. These maps may also be referred to Hover Out of Ground Effect (HOGE) load maps. The main rotor load map 12 generates QROTmain using two input signals, namely, the main rotor speed input signal from the airframe and the collective stick position input from the pilot. The tail rotor load map 14 generates QROTtail using the main rotor speed input from the airframe and the pedal position input from the pilot.

[0022] The aerodynamic (transient) component of the instantaneous rotor torque is more complex than the static components of the instantaneous rotor torque. Therefore, in accordance with a preferred embodiment of the subject invention, it is computed by a polynomial neural network (PNN) using a larger set of sensed input signal from the airframe and pilot. The neural network can be likened to a high-dimensional lookup table. Once is has been created and “trained” the network never changes. It is programmed into the flight control computer of the aircraft in a fixed or “trained” state. Therefore, for a given set of inputs, the output is deterministic and repeatable.

[0023] A polynomial neural network is type of algorithm, and more particularly a self-organizing multi-layered iterative algorithm that can automatically model both linear and non-linear relationships, and yields a polynomial regression equation that is easily interpreted. A PNN algorithm provides robust results in the presence of correlative and irrelative variables or outliers, provides fast learning and numerical stability. PNN algorithms are useful to analyze complex data sets with the aim to determine internal data relationships and to present knowledge about these relationships in the form of a mathematical description.

[0024] Referring once again to FIG. 1, the rotor torque anticipator 10 of the subject invention includes two different forms of a polynomial neural network, which may be selectively engaged through an option switch, depending upon the number of candidate inputs. These networks give a more accurate estimate of the required acceleration/deceleration rate than conventional anticipators because in addition to pilot stick inputs and rotor speed, aircraft body states are taken into account. One network is a Full-Order PNN, also known as a High-Order PNN, the other is a Reduced-Order PNN. A High or Full Order PNN has the advantage of having more information to provide a more accurate prediction of transient torque requirements. A Reduced-Order PNN requires fewer measured parameters, and is therefore less accurate at predicting transient torque requirements, but nonetheless such an algorithm is suitable to provide useful and reliable results in conjunction with the subject invention.

[0025] The Full-Order PNN 16 receives five (5) pilot inputs including collective stick position data, pedal position data, rate of change of pedal position data, lateral stick position data and longitudinal stick position data, as well as four (4) airframe inputs including main rotor speed data, engine torque data, true airspeed data, and pitch attitude data. This data enables the Full-Order PNN 16 to produce the aerodynamic component of the rotor torque QROTaero. In contrast, the Reduced-Order PNN 18 receives three conventional load anticipation inputs including the main rotor set speed, and the collective stick and pedal position data from the pilot, along with the aircraft climb rate. This data enables the Reduced-Order PNN 18 to produce the aerodynamic component of the rotor torque QROTaero.

[0026] In accordance with the subject invention, the two neural networks 16, 18 were trained using the set of maneuvers set forth in Table 1.0. The maneuvers were selected based upon experience with flight testing to size conventional load anticipators.

TABLE 1.0
	Weight
Condition	(lbs)	Atmosphere	Comment
2 sec. Collective Pull	16825	SL/STD	Low Gain Pilot
4 sec. Collective Pull	16825	SL/STD	Low Gain Pilot
8 sec. Collective Pull	16825	SL/STD	Low Gain Pilot
1 sec. Collective Pull	16825	SL/STD	High Gain Pilot
4 sec. Collective Pull	16825	SL/STD	High Gain Pilot
8 sec. Collective Pull	16825	SL/STD	High Gain Pilot
1 sec. Collective Pull	24500	SL/STD	High Gain Pilot
2 sec. Collective Pull	24500	SL/STD	High Gain Pilot
8 sec. Collective Pull	24500	SL/STD	High Gain Pilot
101%/100% split - 1 Sec Collective Pull	16825	SL/STD	High Gain Pilot
101%/100% split - 4 Sec Collective Pull	16825	SL/STD	High Gain Pilot
101%/100% split - 8 Sec Collective Pull	16825	SL/STD	High Gain Pilot
101%/100% split - 1 Sec Collective Pull	16825	SL/STD	Low Gain Pilot
101%/100% split - 2 Sec Collective Pull	16825	SL/STD	Low Gain Pilot
101%/100% split - 8 Sec Collective Pull	16825	SL/STD	Low Gain Pilot
110%/100% split - 2 Sec Collective Pull	24500	SL/STD	High Gain Pilot
110%/100% split - 4 Sec Collective Pull	24500	SL/STD	High Gain Pilot
110%/100% split - 8 Sec Collective Pull	24500	SL/STD	High Gain Pilot
Rapid Hover Turn Right	16825	SL/STD	High Gain Pilot
Rapid Hover Turn Left	16825	SL/STD	High Gain Pilot
High Aggression Roll Reversal	16825	SL/STD	120 kt, 60 deg Roll Right & Left
High Aggression Roll Reversal	16825	SL/STD	120 kt, 60 deg Roll Left & Right
Quick Stop to De-Couple Rotor	16825	SL/STD	120 kt Quick Stop & Pushover
Quick Stop to De-Couple Rotor	16825	SL/STD	80 kt Quick Stop (aggressive)
120 kt Terrain Avoidance (aggressive)	16825	SL/STD	Pull to 2.0 “g”, Pushover to 0 “g”
120 kt Terrain Avoidance (moderate)	16825	SL/STD	Pushover to 0 “g”, Pull to 2.0 “g”

[0027] The Full-Order PNN 16 was designed to predict the aerodynamic component of torque QROTaero in all of the training maneuvers with equal weight placed on each of the maneuvers. That is, the Full-Order PNN 16 is designed to predict aerodynamic torque for collective pitch and non-collective pitch maneuvers. In contrast, the Reduced-Order PNN 18 concentrates primarily on large power transients initiated with collective pitch. The structure of the Reduced-Order PNN 18 is illustrated in FIG. 2.

[0028] Referring to FIG. 2, the Reduced-Order PNN 18 has ten (10) interconnected nodes, with quadratic polynomials stored at each node. These polynomials are set forth hereinbelow. The coefficients of the polynomials are determined during the training phase, wherein numerical optimization techniques are used to curve fit the output of the PNN to test data for each of the training maneuvers. The objective of the training is to minimize the error between the predicted torque and the actual torque versus the time for the combined set of maneuvers. Therefore, there is no single maneuver that is perfectly matched by the PNN. Instead, the best fit for all maneuvers is obtained.

Node1=−47.9016+47.7217*omrm; where omrm is the Main Rotor Speed Ratio (=1 for 100%)

Node2=−0.108098+0.000515121*vc; where vc is the Rate of Climb (Fpm)

Node3=−2.40326+0.0434588*xcpc; where xcpc is the Collective Stick Position (%)

Node4=−3.5805+0.0678252*xppc; where xppc is the Pedal Position (%)

Node5=0.0700913+0.162856*Node2−0.0479842*Node22+0.0301905*Node3+0.120207*Node2*Node3−0.493*Node32−0.074163*Node4−0.589806*Node2*Node4 +0.66802*Node3*Node4+0.100489*Node42.

Node6=−0.154826−0783796*Node1+0.143501*Node12+1.41595*Node3+0.380806*Node1*Node3−0.350551*Node32−0.253151*Node4+0.286963*Node1*Node4−0.216568*Node3*Node4−0.0285866*Node42

Node7=0.381176+0.318371*Node5−0.316399*Node52−0.462027*Node6+0.402343*Node5*Node6−0.156152*Node62−0.0101178*Node2−0.0387345 *Node5*Node2+0.00356685*Node6*Node2−0.0670383*Node22

Node8=0.0937645+1.13228*Node6−0.0208753*Node2+0.269972*Node6*Node2+0.0331041*Node22−1.10191Node3+0.19895*Node6*Node3−0.964847*Node2*Node3+0.174026*Node32

Node9=−48.9181+79.32*Node7

Node10=−4227.53+8130.15*Node8

[0029] The Full-Order PNN 16 is an order of magnitude more complicated than the Reduced-Order PNN 18 because it tries to match collective as well as non-collective maneuvers with equal weighting. Consequently, nine (9) inputs and forty-nine (49) nodes are required in the Full-Order PNN 16.

[0030] Depending upon which neural network is selected, the torque anticipator 10 of the subject invention will sum the QROTaero value from the selected PNN with the QROTmain value generated by the Main Rotor Load Map in Hover at summing junction 20, and this value will be summed with the QROTtail value generated by the Tail Rotor Load Map in Hover at summing junction summing junction 22, resulting in the total instantaneous torque at the rotor blades, QROT. This torque value is representative of the pilot stick inputs, as well as aircraft body states (e.g., airspeed, pitch attitude, roll rate, yaw rate and climb/descent rate), and is used by the engine fuel control system to aid in rotor speed control of the helicopter by delivering an appropriate amount of fuel to the engine to accelerate or decelerate the engine.

[0031] Referring again to FIG. 1, the total torque value QROT is forwarded to the engine control system from the aircraft flight control system in real-time. Initially, the QROT value may be divided in half if the helicopter has two engines. Then, the derivative of QROT is obtained using a differentiator or derivative circuit 24. The resultant is the value rate of change of the total torque QROTDOT (%/sec). This value is conditioned by an amplifier circuit 26 using a conversion factor derived from the gas generator speed. More particularly, an engine performance map 28 is provided which generates a conversion factor based upon the change of gas generator speed NH versus the change is engine shaft torque Q. The conversion factor is applied to QROTDOT to obtain the rate of change of the gas generator speed NHDOT that corresponds to the rate of change of the total rotor torque QROTDOT, or in other words, the engine acceleration/deceleration rate.

[0032] For small values of NHDOT that can be easily handled by a closed loop rotor/power turbine speed governor, the feed forward anticipation logic of the subject invention is not required. Therefore, the value NHDOT is passed through a small deadband 30 to de-couple the load anticipator from the normal engine control logic/laws. This precludes the feed-forward load anticipation of the subject invention from degrading the stability margins of the rotor/power turbine speed control loop.

[0033] By way of example, deadband 30 may be designed to react to signals or values of NHDOT that exceed ±2%/sec. Thus, only steep changes in the rate of change of NHDOT will activate the torque anticipator 10 of the subject invention. The anticipation logic will not react to changes in the rate of change of NHDOT within the predefined deadband range ±2%/sec. Those skilled in the art will readily appreciate that the deadband can be modified to react to a broader or narrower range of signals or values to accommodate differing operating parameters.

[0034] Signals of NHDOT that pass through deadband 30, i.e., those signals or values of NHDOT that fall outside of the deadband range ±2%/sec, represent the demanded rate of change of the gas generator speed NHDOTant. This is the feed-forward anticipation, typically slightly more than NHDOT, to drive the control system onto engine acceleration and deceleration limits. This value is summed with the output of the power turbine governor (PTG) to provide an input to an engine speed controller.

[0035] The engine speed controller manages fuel flow (WF) to the helicopter engine(s) to accelerate or decelerate the helicopter engine(s) in an effort to minimize transient rotor speed droop and overshoot resulting from sudden torque inputs at the rotor blades when the aircraft is maneuvering. At this time, all other anticipators, i.e., collective pitch, rotor speed decay, tail rotor pitch and lateral cyclic anticipators are disabled. Therefore, the only active anticipator is the subject rotor torque anticipator 10. In sum, the subject invention simplifies the engine control system logic by utilizing a single input from the flight control system, namely, the total torque QROT, thus eliminating numerous conventional pilot and airframe inputs.

[0036] Referring to FIGS. 3 through 5, the performance of the neural networks 16, 18 in predicting demanded rotor torque was investigated using three representative helicopter maneuvers that demand large and/or rapid engine power excursions. In each case, the results of the simulation provided a comparison between the following: (a) the true torque demanded during the maneuver; (b) the predicted torque demanded using a full-order polynomial neural network; (c) the predicted torque demanded using a reduced order polynomial neural network; (d) the predicted torque demanded using a reduced order polynomial neural network with 100% NR; and (e) the predicted torque demanded using the steady state Hover Out of Ground Effect (HOGE) Maps for the main and tail rotors.

[0037] Referring to FIG. 3, there is illustrated the plotted results from a simulated recovery from a low-power, rotor coupled decent to a steep climb. The maneuver involved a 2 g collective pull from an 80 knot powered decent to a 25 ft/s climb. In this case, the plotted results show that the Full-Order PNN provides the most accurate prediction of the true torque, as compared to either of the Reduced-Order PNN's. The Reduced-Order PNN's both estimate a somewhat smaller rate of change or torque (about 30% less) than the actual or true torque in the 2-3 seconds time frame. In contrast, the output of the steady state HOGE maps differs substantially from the true demanded torque. This is because transient effects are not taken into account.

[0038] Referring to FIG. 4, there is illustrated the plotted results from a simulated auto-rotation recovery into level flight. The maneuver involved a 2 s collective pull from 10% split needle auto-rotation to level flight. In this case, the static main and tail rotor torque HOGE maps account for only 55% of the change in rotor torque during this maneuver. Therefore, the aerodynamic component of the total torque due to aircraft motion, which is estimated by the PNN, is quite important. A comparison of the two PNN's shows that the Full-Order PNN overshoots the final torque by a considerable amount in the 2.5 to 3.5 second time frame, whereas the Reduced-Order PNN gives a closer match. This is due to the fact that the Full-Order PNN was trained over a wider range of maneuvers, whereas the Reduced-Order PNN was trained primarily on collective maneuvers.

[0039] The Reduced-Order PNN with a constant set speed of 100% NR also does a good job in predicting the true torque. In this case, this configuration is preferable because it effectively de-couples the feed-forward load anticipation from the primary engine control. The bias errors between the predicted torque and the actual or true torque, beyond the 3.5 second mark, are insignificant because the rate of change of torque is small. It should be noted that the load anticipation logic of FIG. 1 takes the derivative of predicted torque to get the engine going which is more important at the start of the maneuver where the rate of change is large. The power turbine governor handles the rest of the maneuver to smoothly control the rotor/power turbine speed.

[0040] Referring to FIG. 5, there is illustrated the plotted results from a high speed terrain avoidance maneuver. In particular, the maneuver involved a 120 knot terrain avoidance pull to 1.5 g, pushover to 0 g. This maneuver uses considerable longitudinal cyclic pitch manipulation, and therefore, it is expected that the Full-Order PNN would provide the most accurate prediction of the true torque. In this instance, the static HOGE maps have little effect on the outcome. The Reduced-Order PNN's do not provide an accurate prediction of torque in this case, because the longitudinal cyclic data is not an input into the Reduced-Order PNN's. Such a shortfall can be accommodated by a responsive power turbine governor. The Full-Order PNN provides an accurate prediction of the true torque because it has the necessary input signals and is trained on similar maneuvers.

[0041] Although the systems and methods of the subject invention has been described with respect to preferred embodiments, those skilled in the art will readily appreciate that changes and modifications may be made thereto without departing from the spirit and scope of the present invention as defined by the appended claims.

Claims

1. A system for determining the total torque required at the main and tail rotors of a helicopter for use in feed-forward rotor torque anticipation comprising: a) means for predicting the aerodynamic torque at the main and tail rotors with the helicopter in motion; b) means for determining the torque at the main rotor with the helicopter stationary; c) means for determining the torque at the tail rotor with the helicopter stationary; and d) means for calculating the required total torque at the main and tail rotors based upon the aerodynamic torque at the main and tail rotors with the helicopter in motion, the torque at the main rotor with the helicopter stationary, and the torque at the tail rotor with the helicopter stationary.

2. A system as recited in claim 1, wherein the means for calculating the required total torque at the main and tail rotors includes means for summing the aerodynamic torque at the main and tail rotors with the helicopter in motion, the torque at the main rotor with the helicopter stationary, and the torque at the tail rotor with the helicopter stationary.

3. A system as recited in claim 1, wherein the means for predicting the aerodynamic torque at the main and tail rotors with the helicopter in motion comprises a polynomial neural network.

4. A system as recited in claim 2, wherein the polynomial neural network is a full order polynomial neural network.

5. A system as recited in claim 2, wherein the polynomial neural network is a reduced order polynomial neural network.

6. A system as recited in claim 4, further comprising means for selecting between the full order polynomial neural network and a reduced order polynomial neural network.

7. A system as recited in claim 1, wherein the means for determining the torque at the main rotor blades with the helicopter stationary is a main rotor load map.

8. A system as recited in claim 7, wherein the main rotor load map is adapted and configured to determine the torque at the main rotor with the helicopter stationary based upon main rotor speed and collective stick position.

9. A system as recited in claim 1, wherein the means for determining the torque at the tail rotor blades with the helicopter stationary is a tail rotor load map.

10. A system as recited in claim 9, wherein the tail rotor load map is adapted and configured to determine the torque at the tail rotor with the helicopter stationary based upon main rotor speed and pedal position.

11. A system as recited in claim 4, wherein the full order polynomial neural network is adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon pilot input data including collective stick position data, pedal position data, rate of change of pedal position data, lateral stick position data and longitudinal stick position data, and airframe input data including main rotor speed data, engine torque data, true airspeed data and pitch attitude data.

12. A system as recited in claim 5, wherein the reduced order polynomial neural network is adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon pilot input data including collective stick position data and pedal position data, and airframe input data including main rotor set speed data and climb rate data.

13. A system as recited in claim 11, wherein the full order polynomial network includes forty-nine interconnected nodes, each node having a quadratic polynomial associated therewith.

14. A system as recited in claim 12, wherein the reduced order polynomial network includes ten interconnected nodes, each node having a quadratic polynomial associated therewith.

15. A system as recited in claim 11, wherein the full order polynomial neural network is trained to predict the aerodynamic torque component of total torque on collective pitch and non-collective pitch maneuvers with equal weighting.

16. A system as recited in claim 12, wherein the reduced order polynomial neural network is trained to predict the aerodynamic torque component of total torque on primarily collective pitch maneuvers.

17. A system for determining the total torque required at the main and tail rotors of a helicopter for use in feed-forward rotor torque anticipation comprising: a) a polynomial neural network adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon a plurality of pilot inputs and airframe inputs; b) a main rotor load map for determining the torque at the main rotor in hover out of ground effects based upon main rotor speed and collective stick position; c) a tail rotor load map for determining the torque at the tail rotor with the helicopter stationary based upon main rotor speed and pedal position; and d) means for calculating the required total torque at the main and tail rotors by summing the outputs of the polynomial neural network, and the main and tail rotor load maps.

18. A system as recited in claim 17, wherein the polynomial neural network is a full order polynomial neural network.

19. A system as recited in claim 17, wherein the polynomial neural network is a reduced order polynomial neural network.

20. A system as recited in claim 18, wherein the full order polynomial neural network is adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon pilot input data including collective stick position data, pedal position data, rate of change of pedal position data, lateral stick position data and longitudinal stick position data, and airframe input data including main rotor speed data, engine torque data, true airspeed data and pitch attitude data.

21. A system as recited in claim 19, wherein the reduced order polynomial neural network is adapted and configured to predict the aerodynamic torque at the main and tail rotors with the helicopter in motion based upon pilot input data including collective stick position data and pedal position data, and airframe input data including main rotor set speed data and climb rate data.