Evolutionary Design and Optimization of Aircraft Engine Controllers
Transkript
Evolutionary Design and Optimization of Aircraft Engine Controllers
554 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 35, NO. 4, NOVEMBER 2005 Evolutionary Design and Optimization of Aircraft Engine Controllers Raj Subbu, Senior Member, IEEE, Kai Goebel, and Dean K. Frederick, Life Member, IEEE Abstract—We present methods to automatically identify and optimize controllers for large-scale complex dynamic systems; in particular, aircraft gas turbine engines. We show how the optimization of different elements within the overall controller can be addressed in an efficient fashion. These elements include local actuator gains, control modifiers, and control schedules. An evolutionary algorithm (EA) is utilized to realize multiobjective optimization on a local as well as a global level, depending on the optimization task at hand. The fitness function comprises performance metrics that incorporate stall margins, exhaust gas temperature, fan-speed tracking error, and local tracking errors. Less attention has been given in the literature to the application of optimization techniques to aircraft engine control systems design, where the controls design and optimization is performed using a full-order engine model and full control systems structures that do not oversimplify the inherent complexities in these highly complex nonlinear dynamic systems. This paper attempts to close that gap. Index Terms—Aircraft engine, automated control, control design, design optimization, evolutionary algorithm (EA), gas turbine. I. INTRODUCTION C ONTROLLERS for modern jet engines have evolved into very complex systems that are difficult to design and require a considerable expenditure of time by many experienced design engineers. Some of the reasons for this situation are the following: • inherent complexity of the engine; • necessity of always protecting the engine and keeping it operating; • importance of efficiency for commercial engines and high performance for military engines. The fact that modern engines are controlled by digital computers allows the designers to implement the high level of complexity necessary to achieve the many and varied operating objectives. However, the combination of a complex engine, varied performance objectives and constraints, and a computer sufficiently powerful for real-time operation of extensive amounts of code results in a formidable design challenge. Manuscript received December 21, 2003; revised August 10, 2004. This work was supported in part by the General Electric Global Research Center under the Automated Controller Design Project. This paper was recommended by Associate Editor Z. Wang. R. Subbu and K. Goebel are with the General Electric Global Research Center, Niskayuna, NY 12309 USA (e-mail: [email protected]; [email protected]). D. K. Frederick is with Saratoga Control Systems, Inc., Saratoga Springs, NY 12866 USA (e-mail: [email protected]). Digital Object Identifier 10.1109/TSMCC.2004.843250 Although current practice makes extensive use of computation for simulation, design calculations, analysis of test results, and verification and validation of performance, the process is very labor intensive and probably falls short of achieving the best possible performance in many cases. Also, the designers who create the final control system as computer software to be executed in real-time on the engine’s electronic control unit (ECU) must be very skilled and have considerable experience in jet-engine control system design. Furthermore, much of their work is done under tight time constraints. Throughout the design process, people are required to make decisions about the organization and structure of the control system and about the numerical values of the many parameters that constitute the thresholds, limits, additive, and multiplicative adjustments, and the tables that implement the many functions of one or two variables that are known as schedules. In current practice, the designers, based on prior experience, computer simulation, satisfaction of constraints, and driven to find the best solution with the resources available, select the parameter values. To achieve this objective, numerous computer simulations are run in an attempt to cover the range of important operating cases. However, the set of cases to be considered is substantial when one accounts for a realistic range of altitudes, Mach numbers, ambient temperatures, different levels of bleeds and power extraction, fuel properties, engine-to-engine quality variations, and deterioration effects. Another consideration is that design work as described previously is best done by engineers with experience, rather than people with a good general controls background but with limited practical experience in the operation and control of jet engines. To deal with these problems, the authors undertook a project to apply computer-optimization tools to some selected aspects of the design of control systems for an in-service high performance commercial engine designed and manufactured by General Electric Aircraft Engines (GEAE). For this example, we assume that the structure of the controller has been determined, including the logic and at least a rudimentary set of schedules that will allow the dynamic response of the engine to be simulated. Furthermore, we assume that the structure of the individual actuator loops has been defined, but that the best values of the gains, as a function of operating conditions, are unknown. Due to the highly nonlinear nature of the engine controller and the fact that it is implemented as a large collection of computer modules (typically over 100) that employ a variety of one- and two-input tables, switching variables, logical elements, limiters, priority-select logic, etc., the control design space is high-dimensional, highly nonlinear, multimodal, and discontinuous. Traditional optimization methods based on nonlinear programming are not likely to be successful in finding an optimal, 1094-6977/$20.00 © 2005 IEEE Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. SUBBU et al.: EVOLUTIONARY DESIGN AND OPTIMIZATION OF AIRCRAFT ENGINE CONTROLLERS or near-optimal solution for such a design problem.1 Another complicating factor governing the selection of an optimization scheme is that a closed-form functional approximation of the performance metric is nontrivial and often impossible. In addition, it is important to define the performance metric in a flexible manner since it is necessary to properly account for such diverse requirements as maintaining stall margins above certain limits, minimizing both peak temperatures and the time spent above a certain temperature, and obtaining short rise times in response to changes in demand values. Furthermore, the minimization must be done over a wide range of flight conditions and disturbance inputs. To accomplish these challenging objectives, we employ evolutionary algorithms (EAs) [1], [15], a class of global stochastic search techniques. Such algorithms execute utilizing principles of natural evolution and are robust adaptive search schemes suitable for searching nonlinear, discontinuous, and high-dimensional spaces. This optimization approach does not require an explicit mathematical description of the objective function over the multidimensional search space, and instead relies solely on an objective function that is capable of producing a relative evaluation of alternative solutions. These features make EAs good candidates for the design of modern jet-engine control systems. The EA in our application is implemented in MATLAB and launches simulation runs of General Electric Aircraft Engine’s controller and engine simulation program, FSIM. II. BACKGROUND Only a very small portion of an overall engine control system is designed to operate in a linear fashion, and even then, the controller gains are often scheduled as functions of the operating conditions (altitude, Mach number, and ambient temperature deviation from standard day). Although much is known about the behavior and design of linear control systems, this information is not relevant to the problems under consideration here. Rather, one must be prepared to work in the nonlinear domain, where theories and analytical results are much more scarce than for the linear domain. Also, the literature on nonlinear control systems, of necessity, tends to deal with specific situations, such as the area of integrator-windup protection (IWP). As indicated in the Introduction, it is not to be expected that conventional optimization methods and those that depend on gradient evaluations, should work. Rather, an evolutionary search algorithm is used which has features and characteristics that make it particularly well suited to the problem at hand. In this section, we present background information on EAs and their application to aircraft engine control systems design. A. EAs EAs include genetic algorithms [11], [13], evolutionary programming [6], [7], evolution strategies [1], and genetic programming [14]. The principles of these related techniques define a general paradigm that is based on a simulation of 1During the initial stages of this technical investigation, we applied a variety of traditional, well-known optimization techniques based on mathematical programming. The performance of these solvers was consistently unsatisfactory, and this was principally because they were ill-suited to the class of problems of our interest. This was a principal factor in pursuing the use of EAs as a robust search methodology. 555 natural evolution. EAs perform their search by maintaining at of any time a population individuals. “Genetic’ operators that model simplified rules of biological evolution are applied to create the new and more . This process continues until a superior population sufficiently good population is achieved, or some other termi, represents via nation condition is satisfied. Each an internal data structure, a potential solution to the original problem. The choice of an appropriate data structure for representing solutions is very much an “art” than “science” due to the plurality of data structures suitable for a given problem. However, the choice of an appropriate representation is often a critical step in a successful application of EAs, and effort is required to select a data structure that is compact, minimally superfluous, and avoids creation of infeasible individuals [15]. For instance, if the problem domain requires finding an optimal real vector from the space defined by dissimilarly bounded real coordinates, it is more appropriate to choose as a representation a real-set-array2 instead of a representation capable of generating bit strings.3 Closely linked to the choice of representation of solutions, , that assigns is the choice of a fitness function credit to candidate solutions. Individuals in a population are assigned fitness values according to some evaluation criterion. Fitness values measure how well individuals represent solutions to the problem. Highly fit individuals are more likely to create offspring by recombination or mutation operations. Weak individuals are less likely to be picked for reproduction, and so they eventually die out. A mutation operator introduces genetic variations in the population by randomly modifying some of the building blocks of individuals. EAs are essentially parallel by design, and at each evolutionary step a breadth search of increasingly optimal subregions of the options space is performed. Evolutionary search is a powerful technique of solving problems, and is applicable to a wide variety of practical problems that are nearly intractable with other conventional optimization techniques. Practical evolutionary search schemes do not guarantee convergence to the global optimum in a predetermined finite time, but they are often capable of finding very good and consistent approximate solutions. However, they are shown (theoretically and practically) to asymptotically converge under mild conditions [16]. B. Evolutionary Design and Optimization of Aircraft Engine Control Systems Evolutionary design and optimization of aircraft engine control systems is a novel applications domain, and therefore there are relatively few reported results in the literature. Multiobjective evolutionary optimization of parameters in simplified engine controllers for simplified Rolls-Royce engine models are reported by Chipperfield and Fleming, and Fonseca and Fleming [3], [4], [8]. Thompson et al. [18] report the selection of components of the controller architecture to meet a range of performance criteria including performance and cost. The choice of architecture is shown to have a large impact on the achievable engine performance. Gremling and Passino [12] 2A real-set-array is an array of bounded sets of reals. representation that generates bit strings can create many infeasible individuals, and is certainly longer than a more compact sequence of reals. 3A Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. 556 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 35, NO. 4, NOVEMBER 2005 report the design of an online adaptive state estimator for a jet engine compressor whose model is evolved by a genetic algorithm. Less attention has been given in general to the application of EAs to aircraft engine control systems design, where the controls design and optimization is performed using a full-order engine model and full control systems structures4 that do not oversimplify the inherent complexities in these highly complex nonlinear dynamic systems. This paper, based on an industrial case-study, attempts to close that gap. III. AUTOMATING CONTROLLER DESIGN Typically, aircraft engine controller design is an iterative process. Initially, a linear engine model is built by extracting partial derivatives from models based on first-principles [5]. Then, local controllers are designed and optimized using first-principles as well as derivatives from previous engine designs. Next, schedules are designed using performance requirements. Finally, the control logic is established which integrates the individual components and takes overall stability and performance requirements into account. The performance of the overall control system is tested on increasingly more complex systems starting with the local model, the bare component level model (CLM), the CLM with the full controller integrated, a dry rig test, wet rig test, test cell runs, and test flight. Each test cycle might necessitate a revision of some controller components with renewed validation and verification. This is a labor-intensive process that will benefit from some level of automation. In particular, design and optimization of the following is beneficial: • local actuator gains, either constant or scheduled; • logic thresholds; • adders and multipliers for gains and schedules; • schedule entries; • control logic structure. This paper describes the optimization for select control variables in the first four categories. Design changes in the control logic structure have not yet been considered. The design and optimization of the controller is decomposed into several complementary subtasks. These subtasks include: 1) optimization of the actuator gains; 2) optimization of the control modifiers (adjustables); and 3) design and optimization of the control schedules. This task decomposition is necessitated by the fact that local gain modifications often do not result in any significant variation at the global performance level. In addition, the potential for crosstalk, i.e., the difficulty to track correlations of several simultaneously manipulated variables on the overall controller, supports the strategy of dividing the optimization endeavor into smaller optimization tasks. Depending on the impact the particular control variable under consideration has on the overall and local performance criteria, we maximize the observability from an optimization standpoint. This means that 4The simulation models we use for the experiments reported in this paper are the very same entities General Electric aircraft engine designers use on a day-to-day basis for actual design and verification prior to embarking on physical module-level verification and validation. Such models are a complex aggregate of a large number of coupled, highly realistic, and field tested simulation entities. Fig. 1. Overall architecture. for some control variables only local performance criteria (local tracking errors) are considered while other control variables are considered from a global level (stall margins, exhaust gas temperature, fan-speed tracking error). Fig. 1 gives an overview of this strategy. The optimization takes advantage of the FSIM simulator that can simulate the dynamic behavior of a production engine and its controller with a high degree of fidelity. The simulation modules comprise the CLM and an emulation of the fully automated digital engine controller (FADEC). The user may specify control settings and flight scenarios, and execute FSIM to obtain the engine response given a high-level (pilot5) command such as demanded fan speed, which is a good measure of thrust. A. Actuator Gains A combination of domain knowledge and several trial runs of FSIM was utilized to isolate and identify the gain candidates to be optimized. Based on the principal consideration of the impact a particular actuator has on engine performance, the fuel metering valve (FMV) proportional gain (FMV-Kp), and the variable stator vane (VSV) proportional gain (VSV-Kp) were selected as parameters to be optimized. While VSV-Kp is a constant, FMV-Kp is a function of the FMV actuator position and the corrected core speed, where a higher core speed results in a higher gain. A simplified view of the proportional path in the VSV actuator control system is shown in Fig. 2, and a simplified view of the proportional path in the FMV actuator control system is shown in Fig. 3. In these actuator loops, d is the demanded actuator position, and p is the achieved actuator position. The optimization problem then is to select that gain value that minimizes the time integral of the square of the position tracking error e of the actuator loop 5The pilot’s thrust specification is via the position of the throttle resolver angle (TRA) measured in degrees, where a lower TRA setting corresponds to a lower thrust demand. Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. SUBBU et al.: EVOLUTIONARY DESIGN AND OPTIMIZATION OF AIRCRAFT ENGINE CONTROLLERS Fig. 2. 557 Proportional path for the VSV actuator. Fig. 4. Control systems architecture. Fig. 3. Proportional path for the FMV actuator. The actuator position demand signal d is computed by the FSIM simulation modules through a complex transformation of the thrust demand profile (with respect to time) specified by the pilot. B. Control System Modifiers/Adjustables The FMV and VSV control systems are large and complex entities with numerous interconnected subsystems. Each of these control systems is provided with a suite of modifiers that consists of adders and multipliers for gains, adders and multipliers for schedules, and logic thresholds (see Fig. 4). In current practice, these control modifiers are selected heuristically. Within the FMV and VSV control systems, a set of over 70 modifiers was identified for optimization. Each adjustable is a bounded real number, and the bounds are specified in FSIM. The optimization problem is to identify a set of modifiers such that global performance criteria are optimized. Ideally, the performance metric should be aggregated over a number of flight conditions such as altitude, Mach number, and ambient temperature deviation from standard day and engine configurations such as horsepower extraction, deterioration, and component tolerances. To adequately evaluate performance, a metric is necessary that allows the quantification of all global performance requirements. Such an ideal global performance metric should include the following relative measures: • booster stall margin versus booster inlet flow; • compressor stall margin versus compressor inlet flow; • VSV demanded position versus corrected core speed; • Variable bleed valve (VBV) demanded position versus corrected fan speed; • corrected Phi6 versus corrected core speed; • combustor fuel/air ratio versus any related severity parameters; • high-pressure turbine inlet temperature versus corrected core speed; • exhaust gas temperature versus corrected fan speed and time. To reduce the complexity of the global performance metric, we focus on typical input variations and study the most critical 6Phi is the ratio of the fuel flow and the corrected combustor static pressure. parameters for aircraft engine control systems validation. In particular, we require meeting all stall margin limits, good tracking of a fan-speed demand profile, and reduction in the peak exhaust gas temperature. Toward this, let EGT be the the acceptable cruise exhaust gas temperature profile, EGT temperature, the exponent by which the distance to the limit is penalized, a weight for the temperature component, a weight the fan-speed for the fan-speed tracking error component, the fan-speed demand profile, the time, E the profile, exceedance profile comprising the EGT exceedance , the , the booster stall margin fan stall margin exceedance , and the compressor stall margin exceedance exceedance . Then the optimization problem has the form , where EGT EGT and if EGT EGT otherwise if SM SM otherwise if SM SM otherwise if SM SM otherwise C. Schedules The control logic in a typical aircraft engine controller utilizes a suite of schedules that are functions of one or two input variables. Schedules are implemented as lookup tables and the output values are computed via linear interpolation among the closest neighbors. Schedule surfaces represent nonlinear transformations of the inputs to the output and are critical components of an aircraft engine’s control logic. Based on a combination of domain knowledge, simulation, and knowledge elicitation from domain experts, a particular schedule in the FMV module was selected as a candidate for Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. 558 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 35, NO. 4, NOVEMBER 2005 Fig. 5. Optimization of VSV-Kp for burst-at-cruise. evolutionary optimization. This schedule outputs a rate-gain reduction given the ambient pressure (a physical function of altitude) and compressor speed, and is active during the burst phase for a specific maneuver called a Bodie, wherein at cruise the pilot cuts thrust for a short time period and increases thrust through a burst before the engine temperatures have achieved steady state at the reduced power level. In the absence of a rate-gain schedule, the fan-speed response during the burst phase is extremely sluggish, which is highly undesirable. What is desirable however, is a rapid return to the original fan speed. The optimization problem is to identify a set of schedule entries such that fan acceleration is maximized during the burst phase of the Bodie, subject to maintaining all stall margins above acceptable limits. Simulation-based evaluation reveals that during this maneuver the EGT is always well within limits, so it is not included in the global performance metric. Let be the fan-speed profile, the fan-speed demand profile described as a step function with the step coinciding with the burst phase of the Bodie, the time, E the exceedance profile , the booster comprising the fan stall margin exceedance , and the compressor stall margin stall margin exceedance . Then the optimization problem has the exceedance , where form and if SM SM otherwise if SM SM otherwise if SM SM otherwise An evolutionary optimization of a schedule that is the function of two input variables corresponds to a systematic and joint manipulation of the table entries. An important requirement in the automatic design and optimization of these control surfaces is the smoothness of these derived surfaces. Unless smoothness is explicitly included as a design requirement, an evolutionary optimization can result in noisy albeit optimal, schedules. To facilitate smoothness in derived schedule surfaces, the entries in each test surface T are filtered using a specialized bi-directional , filtering algorithm that is applied to each derived row and is shown in the following: 1) 0 Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. SUBBU et al.: EVOLUTIONARY DESIGN AND OPTIMIZATION OF AIRCRAFT ENGINE CONTROLLERS Fig. 6. Optimization of VSV-Kp for Bodie-at-cruise. Fig. 7. Optimization of FMV-Kp for burst-at-cruise. 2) 3) 559 (line 3) for each row following the procedure outlined in lines 1, 2, and 3. This is an important step, since selection of a reliable starting point is critical. Next, the smoothed are computed using the procedure outlined in values line 4. IV. RESULTS 4) In the previous algorithm, is a smoothing factor. The objective of the algorithm is to first identify a reliable starting value This section describes the initial observations followed by systematic experiments that led to the simulation-based optimization of actuator gains, controls modifiers, and schedules. The ambient conditions (e.g., altitude) selected for these experiments correspond to typical operational regimes of interest Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. 560 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 35, NO. 4, NOVEMBER 2005 Fig. 8. Optimization of FMV-Kp for Bodie-at-cruise. Fig. 9. Optimization of FMV control modifiers, response for burst-at-cruise. to engine designers. While we explored a variety of ambient conditions, due to space considerations, we only report a select representative sample of these results. A. Preliminary Observations In selecting candidates to which the search algorithm was to be applied, several considerations were taken into account. First, it was decided to start in an area of constrained complexity and about which enough knowledge was accessible. The first subject was the code that handles the adjustments to the VSV positions. This operation can be divided into two parts: 1) the determination of the demanded VSV position, based on current flight conditions and 2) the gains and other parameters of the VSV actuator loop itself. Upon looking at the Beacon7 diagrams that describe the code used in the ECU of the engine, it was concluded that the gains of the actuator loop made a better starting point than the demand-value determination. For the latter, 12 control modules, implemented by 12 different Beacon diagrams, are involved. Also, without an intimate knowledge of how these 12 modules operated, it would be difficult to pose and carry out a meaningful optimization problem. Hence, it was decided to start by focusing attention on the tuning of the VSV actuator loop. Specifically, experiments were carried out with the proportional gain of that loop, since the code had an adjustable adder defined which could be set to any desired value at the start of a run, without having to rebuild the FSIM code. 7Beacon is the computer program used by GE Aircraft Engines to represent the controller in terms of block diagrams and computational flow diagrams from which the actual computer code is generated. Applied Dynamics, Inc., Ann Arbor, MI, markets it. Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. SUBBU et al.: EVOLUTIONARY DESIGN AND OPTIMIZATION OF AIRCRAFT ENGINE CONTROLLERS Fig. 10. Optimization of FMV control modifiers, response for Bodie-at-cruise. Fig. 11. Optimization of VSV control modifiers, response for burst-at-cruise. B. Actuator Gains 1) Runs With Preset Values of VSV-Kp: Several FSIM runs were made for a combination of a burst (increase of TRA from 36–78 ), a constant TRA for 25 s, followed by a chop (decrease in TRA from 78–36 ). For a given FSIM run, the value of the proportional-gain adder was varied such that the actual proportional gain was a constant between the limits of 78 and 778. The design value is a constant (not scheduled according to core speed or any other variable). In terms of the key engine variables, such as fuel flow, fan speed, VSV angle, exhaust gas temperature (EGT), etc., there were virtually no differences over the tenfold range of the proportional gain. The only variable that was affected by the change in proportional-gain value was the error in the VSV actuator loop. This observed behavior suggests 561 that the proportional gain could be selected so as to minimize the square of the actuator error. 2) Optimization of VSV-Kp: Optimization of the VSV-Kp follows the procedure outlined in Section III-A. Engine operation was simulated subject to changes in throttle position while cruising at 35 000 ft, Mach 0.8, and standard-day temperature. A burst and Bodie are used to excite the overall system, and gain optimization is performed independently for each of these excitation profiles. Fig. 5 shows the results of optimization of VSV-Kp for the burst, while Fig. 6 shows the results for VSV-Kp optimization under a Bodie maneuver. In each of these cases the optimized gain is substantially higher than the default gain, and this results in a significant reduction of the position tracking error over the space of the excitation. However, while there is a noticeable impact due to optimization Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. 562 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 35, NO. 4, NOVEMBER 2005 Fig. 12. Optimization of VSV control modifiers, response for Bodie-at-cruise. on the actuator level performance, there is no global impact due to optimization. A key observation is that VSV-Kp being a gain that is not scheduled has different optimal values for the burst maneuver and the Bodie maneuver. Although this was an interesting and useful first step, this result is probably of only limited significance. No attempt was made to determine optimized values for other gains in the loop, such as the integral gain, or other parameter values, such as a lead-time constant. However, it would certainly be possible to do so. 3) Runs With Constant Values of FMV-Kp: Next, this actuator design method was applied to the FMV actuator loop proportional gain. Here, the actuator controller is slightly more complicated than the VSV one in that the proportional gain is a function of both the FMV actuator position and the core speed, where a higher core speed results in a higher gain. As in the VSV actuator study, it was found that changes in the FMV proportional gain had virtually no effect on the engine variables (fan speed, fuel flow, temperatures, etc.). The scheduling of the gain was removed by modifying the appropriate schedule table and the burst and chop simulations were run over a wide range of constant gains. As before, only the error in the actuator loop was affected. 4) Optimization of FMV-Kp: Optimization of the FMV-Kp follows the procedure outlined in Section III-A. Engine operation was simulated subject to changes in throttle position while cruising at 35 000 ft, Mach 0.8, and standard-day temperature. A burst and Bodie are used to excite the overall system, and gain optimization is performed independently for each of these excitation profiles. Fig. 7 shows the results of optimization of FMV-Kp for the burst, while Fig. 8 shows the results for FMV-Kp optimization under a Bodie. As observed from the optimization of the VSV-Kp, in each of these cases, the optimized gain is substantially higher than the default gain, and this results in a significant reduction of the position tracking error over the space of the excitation. In order to find critical parameters whose values will affect global performance metrics (such as stall margins or exhaust gas temperature), attention was focused on the extensive set of ECU modules that are used to produce the incremental changes in the demanded fuel flow. These incremental changes are continuously summed in order to produce the demand value for the FMV actuator. The behavior during a burst operation, in response to a sudden request from the pilot for an increase in power, was examined. During such a situation, several different regulators are active, depending on the various limits and constraints that must be satisfied in order to guarantee safe operation of the engine. By looking at which regulator was selected as time evolved following the burst command, we determined the sequence of active regulators. The study of control logic charts (“Beacon diagrams”) for the modules that are active during the flight maneuver of interest gave understanding of their operation and the conditions that cause a transition from one module to another. This allowed the formulation of an optimization problem that did produce some interesting results for a problem whose complexity is comparable to what would be required to do in a real design. The optimization task focused then on the modifiers and schedules as described in the following. C. Optimization of Control System Modifiers Determining the values for the control modifiers in the FMV and VSV control systems follows the procedure outlined in Section III-B. Engine operation was simulated subject to changes in throttle position while cruising at 36 000 ft, Mach 0.8, and standard-day temperature. Fig. 9 shows the results of optimization of the FMV control system modifiers under a burst-at-cruise excitation, and Fig. 10 shows the result of optimization of the FMV control system modifiers under a Bodie-at-cruise excitation. Unlike the procedure followed for optimization of actuator gains, where optimization was performed independently for each of the excitation profiles, the optimization of control modifiers for each control system is performed jointly over the two excitation profiles. Such an approach helps determine the best set of control modifiers8 for the two excitation profiles combined. In practice, such an approach could be extended to include other excitation profiles corresponding to other flight scenarios of interest. It is seen that the FMV control modifiers have a significant impact on engine performance at the global level, and optimization results in a solution set that significantly 8The normalized form of each modifier, default, and optimized, is presented for purposes of comparison. For each normalized modifier, 0 represents its minimum allowable value and 1 represents its maximum allowable value. Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. SUBBU et al.: EVOLUTIONARY DESIGN AND OPTIMIZATION OF AIRCRAFT ENGINE CONTROLLERS 563 reduces exhaust gas temperature transients albeit at the slight expense of fan acceleration during a burst-at-cruise. However for the Bodie-at-cruise, the exhaust gas temperature transients are significantly reduced and fan acceleration is improved. Moreover, the stall-margin limits are strictly followed. Figs. 11 and 12 show corresponding results of optimization of the VSV control system modifiers. While there is a slight reduction due to optimization in the exhaust gas temperature transient, there is no perceivable impact on the fan acceleration, which is directly related to engine thrust. This observation lends experimental evidence to the common knowledge that considered independently, certain control system modules can have a more significant impact than others at the global performance level. D. Optimization of Schedules Optimization of the schedule rate-gain entries follows the procedure outlined in Section III-C. Engine operation was simulated subject to changes in throttle position while cruising at 41 000 ft, Mach 0.8, and standard-day temperature. Fig. 13 shows the default and optimized rate-gain schedule surface, where the optimization elevates the rate gain over the input space in order to improve fan acceleration. Fig. 14 shows the performance metrics that exhibit differences due to the default and optimized schedules during the Bodie maneuver. The key observations are that schedule optimization results in a substantial improvement in fan acceleration without sacrificing exhaust gas temperature limits and stall margin limits. Also, the minimum stall margins with the optimized table are no lower than those with the default table. E. Discussion In optimization applications involving EAs, a principal issue is the execution speed of the underlying models and its impact on the speed, and consequent quality of convergence achievable given computational resources. Though we used highly realistic and complex engine simulation modules, the time to execute one complete run of the simulation modules was of the order of a few seconds on typical desktop processors. Therefore, a typical evolutionary search did not take more than a few hours to complete. In realistic design scenarios, it would be highly advantageous and attractive to designers if they were to get near instantaneous feedback within a few minutes at most. For such scenarios, it would be suitable to distribute the simulation-based evaluation tasks to multiple processors to meet the optimization response-time requirements. Determining values of the controls modifiers so as to simultaneously track a reference fan-speed profile and reduce the exhaust gas temperature, while adhering to stall margins and temperature limits, is a multiobjective optimization problem. The issue of the manner in which the individual terms in the objective function are weighted is critical, because the optimization method will try to satisfy the objective function depending on the weights placed on individual criteria. If one criterion is weighted heavily relative to the others, the optimization result will be skewed in favor of this particular criterion. Therefore, a careful balance that reflects what is really important needs to be found. Fig. 13. Default and optimized schedule surfaces. Another issue is the design of the individual criteria. Although a general idea exists as to what the optimization ought to accomplish, there are differences in the implementation of the objectives. For example, the optimizer could be asked to closely follow an ideal fan-speed profile. The question then arises as to how such an ideal fan-speed profile should look. It is necessary to draw upon domain knowledge to properly trade off the different criteria and to properly set up the objective function. Finally, the results of a multiobjective optimization could be expressed as a Pareto frontier [2], [3], [4], [8]. In other words, there are a number of possible solutions that all meet the criteria that may result in different engine behavior. It is then necessary to express a preference for a particular solution from the set of possible solutions. Ideally, the preference could be cast within the objective function. However, the effects of the solutions are not always apparent until after the optimization results are reviewed. V. SUMMARY AND CONCLUSION We have formulated a framework to the design of jet-engine controllers by applying evolutionary search algorithms to actuators, multiplicative, and additive adjustments, and table generation and/or modification. To that end, we developed meaningful performance functions whose minimization produces controller parameters that result in desirable engine behavior. It incorporates stall margins, EGT, and tracking of changing throttle positions. In addition, we employed a smoothing function that in effect penalizes discontinuous table solutions. We then used these techniques successfully to adjust over 70 parameters at a time in the controller modules of a real commercial engine. In addition, we demonstrated the ability to tune the proportional gain of a regulator in the context of its operation in a nonlinear environment by minimizing the integral of the square of the actuator error. Finally, we employed evolutionary search algorithm methodology to generate a 3-D table of the form to maintain rapid response to a demanded power increase as a part of a Bodie maneuver. Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. 564 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 35, NO. 4, NOVEMBER 2005 Fig. 14. Schedule optimization results. There are a number of avenues for future work. One avenue is the integration of the optimization for design assistance during the various design and validation stages—cycle deck over CLM, FSIM, dry rig, wet rig, test cell, and flight test. Moreover, this optimization approach may be extended to adapt engine performance based on in-service data, and to adapt engine performance as an engine deteriorates. This assistance could range from automated validation of design choices to suggestion of parameters as shown in this paper. Another avenue leads to scaling the optimization task. Of particular interest is ensuring cross-communication of individual results from components in a concurrent optimization scheme. Such a coevolutionary optimization [17] would allow concurrent module-specific exploration of the global design space, thus, responding to the need of both domain-specific focus and adhering to global performance metrics. This could be accomplished via agent-based multiobjective optimization. Also of interest is the integration of external information such as expert knowledge, historical runs, information from pilots during test flights, etc. This task would need the development of an information aggregation component [9], [10] that can deal with the inherent uncertainties and the different format to more formally translate these observations into an objective function and performance metric. Another avenue could lead to a fault-tolerant controller that would respond to a fault signature with appropriate changes to the control structure. This could be an extension of the compensatory optimal control using quality and deterioration estimates. Similarly, one could perform individualized optimization of engines using modifiers by responding to specific engine characteristics (as opposed to model wide baselines), thus, further improving performance. In addition, one could also explore performance-enhancing optimization through the reduction of schedule size, thus, reducing FADEC memory requirements and improving throughput. This could lead to selection of optimal schedule sizes for a number of controllers such as the FMV and power management, which typically deal with large schedules. In addition, the overall FADEC architecture could be optimized by identifying obsolete elements (schedules, etc.). Finally, we mentioned as one opportunity for optimization the logic structure itself. This could be accomplished through genetic programming or inductive learning such as experience based learning. ACKNOWLEDGMENT The authors would like to thank General Electric for helpful cooperation during the preparation of this manuscript. They would also like to thank the reviewers for their efforts, and for their very helpful suggestions. REFERENCES [1] T. Bäck, Evolutionary Algorithms in Theory and Practice. New York: Oxford Univ. Press, 1996. [2] S. Bonissone and R. Subbu, “Exploring the pareto frontier using multisexual evolutionary algorithms: An application to a flexible manufacturing problem,” in Proc. SPIE Annu. Meeting—Applications and Science of Neural Networks, Fuzzy Systems, and Evolutionary Computation V, vol. 4787, 2002, pp. 10–22. [3] A. Chipperfield and P. Fleming, “Multiobjective gas turbine engine controller design using genetic algorithms,” IEEE Trans. Ind. Electron., vol. 43, no. 5, pp. 583–587, Oct. 1996. [4] A. J. Chipperfield and P. J. Fleming, “Evolutionary design of gas turbine aero-engine controllers,” in Proc. IEEE Int. Conf. Systems, Man, and Cybernetics, 1998, pp. 2401–2406. [5] C. M. Close, D. K. Frederick, and J. Newell, Modeling and Analysis of Dynamic Systems, 3rd ed. New York: Wiley, 2001. [6] D. B. Fogel, Evolutionary Computation: Toward a New Philosophy of Machine Intelligence. Piscataway, NJ: IEEE Press, 1995. [7] L. J. Fogel, A. J. Owens, and M. J. Walsh, Artificial Intelligence Through Simulated Evolution. New York: Wiley, 1966. [8] C. M. Fonseca and P. J. Fleming, “Multiobjective optimization and multiple constraint handling with evolutionary algorithms—Part II: Application example,” IEEE Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 28, no. 1, pp. 38–47, Jan. 1998. Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply. SUBBU et al.: EVOLUTIONARY DESIGN AND OPTIMIZATION OF AIRCRAFT ENGINE CONTROLLERS [9] K. F. Goebel, M. Krok, and H. Sutherland, “Diagnostic information fusion: Requirements flowdown and interface issues,” in Proc. IEEE Aerospace Conf.—Advanced Reasoner and Information Fusion Technique, 2000, pp. 155–162. [10] K. F. Goebel, “Architecture and design of a diagnostic information fusion tool,” Artif. Intell. Eng. Des., Anal., Manufact., vol. 15, no. 4, pp. 335–348, Sep. 2001. [11] D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning. Reading, MA: Addison-Wesley, 1989. [12] J. R. Gremling and K. M. Passino, “Genetic adaptive state estimation for a jet engine compressor,” in Proc. 12th IEEE Int. Symp. Intelligent Control, 1997, pp. 131–136. [13] J. H. Holland, Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial Intelligence, 3rd ed. Cambridge, MA: MIT Press, 1994. [14] J. Koza, Genetic Programming: On the Programming of Computers by Means of Natural Selection. Cambridge, MA: MIT Press, 1992. Data Structures [15] Z. Michalewicz, Genetic Algorithms Evolution Programs. New York: Springer-Verlag, 1996. [16] R. Subbu and A. C. Sanderson, “Modeling and convergence analysis of distributed coevolutionary algorithms,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 34, no. 2, pp. 806–822, Apr. 2004. , “Network-based distributed planning using coevolutionary [17] agents: Architecture and evaluation,” IEEE Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 34, no. 2, pp. 257–269, Mar. 2004. [18] H. A. Thompson, P. J. Fleming, A. J. Chipperfield, and C. Legge, “Distributed aero-engine control systems architecture selection using multiobjective optimization,” Control Eng. Pract., vol. 7, no. 5, 1999. + 565 Kai Goebel received the Ph.D. degree in mechanical engineering from the University of California at Berkeley. He is a Senior Research Scientist in the Computing and Decision Sciences Group, General Electric Global Research Center, Niskayuna, NY. He is also an Adjunct Professor in the Computer Science Department, Rensselaer Polytechnic Institute (RPI), Troy, NY, since 1998, where he teaches classes in soft computing. He has carried out applied research in the areas of artificial intelligence, soft computing, and information fusion. He has worked on using soft computing techniques for real time monitoring, diagnosis, and prognosis of industrial equipment such as aircraft engine, and power plants, and structures such as aircraft wiring. He has also carried out research for both data fusion and decision fusion in mechanical systems as well as financial applications. He has published more than 50 articles, including two book chapters. He holds five patents. = Raj Subbu (M’00–SM’04) received the Ph.D. degree in computer and systems engineering from Rensselaer Polytechnic Institute (RPI), Troy, NY, in 2000. Since 2001, he has been a Senior Research Scientist in the Computing and Decision Sciences Group, General Electric Global Research Center, Niskayuna, NY. His research interests are in the areas of information systems, control systems, novel multiobjective optimization methodologies, and soft computing. At General Electric, he serves as a principal technologist at the intersection of these areas for several high-business-impact projects. In addition, he served as Co-Principal Investigator in a multiyear (2001–2004) National Science Foundation funded project in scalable enterprise decision systems, in collaboration with RPI. He has authored over 25 publications and proceedings, has received one U.S. patent, and has 15 U.S. patents pending. He is the principal coauthor of the book Network-Based Distributed Planning Using Coevolutionary Algorithms (Singapore: World Scientific). Dr. Subbu received the Best Paper Award at the IEEE International Conference on Fuzzy Systems in 2003. He is an Associate Editor of the IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS—PARTS A and C. Dean K. Frederick (S’60–M’64–LM’00) received the B.E. degree in mechanical engineering from Yale University, New Haven, CT, the Sc.M. degree in electrical engineering from Brown University, Providence, RI, and the Ph.D. degree in electrical engineering from Stanford University, Stanford, CA. After teaching at Clarkson University, Potsdam, NY, he spent 30 years as a Faculty Member in the Electrical, Computer, and Systems Engineering Department, Rensselaer Polytechnic Institute (RPI), Troy, NY. During this period, he coauthored more than 40 papers in the area of modeling, simulation, automatic control, and computer-aided control system design. He also coauthored two text books dealing with linear systems and the modeling and analysis of dynamic systems. He had a number of visiting positions, including NASA’s Marshall Space Flight Center, Huntsville, AL; the Exxon Corporation, Linden, NJ; General Electric’s Corporate Research Center, Niskayuna, NY; The Control Systems Center, UMIST, Manchester, U. K.; Lund Technical Institute, Lund, Sweden; and Technical University of Delft, The Netherlands. He also consulted for General Electric Company, Emhart Glass, Lawrence Livermore National Laboratory, and The Applied Physics Laboratory of The Johns Hopkins University. After retiring from Rensselaer in 1994, he spent six years as a senior engineer with Unified Technologies, Troy, NY, where he did contract work related to the control of jet engines for General Electric’s Aircraft Engines Division and their Global Research Center. Areas of work included multivariable control, simulation, fault detection, and the development of graphical-user-interface (GUI) systems for the analysis and control of both military and commercial jet engines. In June 2002, He established his own company, named Saratoga Control Systems, in Saratoga Springs, NY. He continued performing contract work for GE until September 2003, and also did an engine-related project for RLW, Inc. of State College, PA. In December 2003, he began work for the Glenn Research Center of NASA on a project to produce a Simulink dynamic model of a commercial jet engine. At present, he is also working on advanced power-plant controls with GE’s Global Research Center in Niskayuna, NY. He is a coauthor of two MATLAB books, dealing with continuous-time and discrete-time control systems. Dr. Frederick is a member of the IEEE Control Systems Society. He received a Centennial Certificate from the American Society of Engineering Education for his work with computers in engineering education. Authorized licensed use limited to: CZECH TECHNICAL UNIVERSITY. Downloaded on October 5, 2009 at 07:23 from IEEE Xplore. Restrictions apply.