Intelligent and Autonomous Flight Technology for Launch Vehicles

SONG Zhengyu

China Academy of Launch Vehicle Technology, Beijing 100076

Abstract: This paper proposes the architecture of an intelligent flight launcher system as well as fundamental solutions to capability prediction and dynamic planning. This effort reflects the latest progress in the applications of intelligent and autonomous technology for launcher flights. The paper first describes the characteristics and capabilities of intelligent and autonomous systems and classifies various related technologies. In the context of intelligent and autonomous technology in aerospace engineering, it then focuses on technical difficulties involved with intelligent flight and reviews developments in the field. An E3 classification model of an intelligent flight launcher is then proposed and its application scenarios are discussed. Based on an intelligent flight system configuration of the launcher, the online trajectory planning and initial value guess are examined, and vertical landing is provided as an example to explain the effects of the implementation of computational intelligence to flight systems.

Key words: intelligent flight, launch vehicle, autonomous, hierarchical model, dynamic trajectory planning

1 BASIC CONCEPT OF INTELLIGENT AND AUTONOMOUS SYSTEMS

1.1 The Capabilities of Intelligent and Autonomous Systems

An accurate definition of “intelligence” or “autonomy” is not the focus of this paper. The meanings of these two concepts are determined based on the performance, or capability, of intelligent and autonomous systems.

“Intelligence” generally refers to the manifestation of highorder brain activity. For unmanned systems, it is the capability of automatically acquiring and applying knowledge, thinking,reasoning, problem solving, and learning - in order words, a comprehensive ability to analyze, judge, act, and deal with issues effectively in a given environment. “Autonomy” refers to the ability to perceive, observe, analyze, communicate, plan,make decisions, and act, and reveals the state and quality of selfmanagement of unmanned systems.

An intelligent and autonomous system has a high level of autonomy and intelligence which is reflected in the following aspects[1]:

Perception: This is a basic concept in cognitive psychology,and includes sensation and consciousness. Sensation is the reflection of the attributes of objective things and consciousness the overall reflection thereof.

Understanding: This involves a more complex information processing capability than perception; it is more abstract and subjective.

Motion: This is the ability to change spatial location, and involves such notions as rapidity, agility, and maneuverability.

Operation: This is the ability to change the external environ-ment to achieve a certain purpose.

Both motion and operation need to be realized through control.

Learning: This is the ability to process and refine external information to form knowledge. Training is a common method,and autonomous and transfer learning are subjects of intensive research at present.

Adaptation: This is the ability of a system to actively adjust or change its parameters, state, structure and behavioral characteristics to adapt to changes in the given task, environment, and its own internal state.

Planning: This is the process of calculating a series of orderly actions to change the system from the given to the expected state by allocating resources, meeting and optimizing various constraints.

Decision making: This is the process of cognition used to choose from among various alternatives.

Communication: This is spontaneous, active, and purposeful transmission of information to other subjects.

Coordination: This is the process of multiple subjects jointly accomplishing a common goal by emphasizing overall benefit and sacrificing personal benefit if needed.

Communication and coordination are behaviors exhibited by intelligent and autonomous systems interacting with other subjects.

Intelligent and autonomous flight involves advanced Artificial Intelligence (AI), and is a challenging research area.

1.2 Classification of Intelligent and Autonomous Technology

Technologies related to intelligent and autonomous systems can be categorized into the following five aspects[2]:

1) Agent-based reasoning, such as in rule-based expert systems, Bayesian belief networks, particle filtering,case-based reasoning, and fuzzy logic.

2) Biologically inspired reasoning, including neural network(NN), genetic algorithm (GA), ant colony optimization,and the artificial immune algorithm.

3) Machine learning systems, including data mining, supervised and unsupervised classifiers, and deep NNs.

4) Naturalistic interface, including natural language processing, semantic analysis, and speech and gesture recognition.

5) Hybrid modeling.

Not all the above technologies can be used in flight, and only those relevant to flight are discussed below.

2 STATS OF RESEARCH ON INTELLIGENT AND AUTONOMOUS FLIGHT TECHNOLOGY

2.1 Intelligent Systems for Aerospace Engineering

Intelligent systems can be widely used in aerospace engineering[3]and their most common application is in the modeling of mathematical equations, rule-based systems, fuzzy models,neural models, and tree structures. In optimization problems,intelligent systems can help search strategies, such as depth-first search, heuristic dynamic programming, and GAs. They can also assist in design, training, and such other areas as decision making, planning and scheduling, health management, prediction,and knowledge discovery. However, few have been used realtime in space vehicles.

The characteristics of intelligent and autonomous systems are difficult to quantify. From the concepts discussed above, it is evident that the performance of intelligence and autonomous systems in space flight is closely related to control technology,and there is no clear boundary between intelligent and unintelligent, and autonomous and non-autonomous, systems.

Table 1 lists the definitions of ESA for the hierarchical levels of an autonomous system[4].

Many aerospace scientists and engineers prefer to refer to machine learning and autonomy rather than AI. No industry classification or agency standard has thus far been formulated for intelligent flight. The academic perspectives provided in Table 2 are for reference only[5].

2.2 Challenges facing GNC

Intelligent and autonomous flight poses significant challenges to GNC systems. Autonomous mission planning and onboard trajectory design in particular correspond to a higher level of intelligence and autonomy, which enhances safety and chances of the success of the mission. The roles and challenges of, and the impediments in these areas are listed in Table 3[6].

Trajectory planning is a fundamental task for aerospace engineers that is usually conducted offline before flight. Reference tracking-based guidance control is used in this situation, but it cannot deal well with uncertainty. Future exploration missions will require increasing independence from ground control, because of which online or dynamic planning, representing high-level intelligence and autonomy, will be needed using the following approaches[7]:

Table 1 Levels of autonomy: A system capable of fulfilling its purpose without external intervention

Table 2 Levels of intelligent control: The capabilities for self-improvement

Table 3 Challenges of autonomous planning / guidance

· automatic and smart initial guess generators;

· powerful mission design and optimization software; and

· new technologies to support real-time onboard operations for new flight regimes.

Some basic methods, or automatic tuners, provide support for these approaches, and include numerical integration, array gradient calculations, and intelligent decision algorithms to guide and redirect complex optimization sequences.

2.3 Results of Prevalent Research

In Ref.[8], the human consciousness was mapped to a flight control system where the control/estimation compensation,the actuators, was considered a reflexive function processed by bypassing the brain. Navigation, adaption, and guidance were mapped to the human subconscious and classified as procedural functions. System identification was assumed to be similar to the pre-conscious while system monitoring and goal planning were assumed to correspond to the conscious, and were both regarded as declarative functions. Through this mapping, an intelligent and autonomous control system with bionic features was expressed.

Reflexive functions and NN-based algorithms are widely used in the inner loops of control systems (attitude control) to deal with uncertainties. Many algorithms have been developed for scenarios that involve controlling a damaged aircraft and improving performance with large uncertainties[9], such as the F-15 intelligent flight control system (dynamic inversion and online learning network) developed by NASA[10]; adaptive control for open-loop unstable systems developed by the USAF’s RESTORE; and robust adaptive control for very flexible aircraft(observer-based control with loop transfer recovery, OBLTR)by Boeing[11].

Outer loops, such as mission planning and management,require interaction with a human operator, and breakthroughs in terms of system autonomy in these areas have been relatively few. With better computational performance, numerical calculation-based methods have garnered research interest and convex optimization has commonly used to solve global optimization problems. Guidance for fuel-optimal large diverts (G-FOLD)[12,13], led by the JPL and demonstrated in the ADAPT project[14,15], has been publicly reported. The project involves assessing a guidance method based on convex optimization for the powered descent phase of a Mars landing to test its capability for optimal, large-scale maneuvering. The G-FOLD algorithm is regarded as the fuel-optimal method of autonomous guidance but has not been verified on recent Mars missions.

3 VIEWS ON INTELLIGENT AND AUTONOMOUS FLIGHT TECHNOLOGY FOR LAUNCHERS

3.1 E3 Hierarchical Model

The E3hierarchical model is shown in Figure1, and represents the levels of intelligent and autonomous flight of launchers.

Figure 1 E3 hierarchical model

Level 0: Clear problem and model, and easy to solve. The main challenge is uncertainty, which corresponds to the traditional scenario of tracking guidance and attitude control. In modern control theories, adaptive control, ARDC, and robust control are applied to solve uncertainty. For intelligent solutions,biology-inspired reasoning is widely used.

Level 1: Clear problem and an ideal model, but difficult to solve. If the model can be simplified, this level may degrade to Level0, but this simplification often requires compensation at the cost of other subsystems. Mission or trajectory planning is part of this layer, and requires experience. The solution is dependent on a growing extent with the computer’s performance.The concept of computational intelligence has thus been proposed because the process reflects a certain degree of intelligence through calculation. However, a higher computational efficiency does not guarantee convergence or the optimal solution, and initial value generation has gradually become the focus of research in this area. Smart methods have been considered to determine the initial value, which is the meaning of intelligent computation in this context.

Level 2: Unclear problem and little experience. The “eureka”moment, or human inspiration, is implied. At present, this has no suitable machine-based solutions.

3.2 Demand for Intelligent and Autonomous Flight

Considering the flight profiles of launchers, intelligent and autonomous technology function effectively under the following conditions:

1) A transition is required from serial or sequential offline optimization to global, “end-to-end”, and real-time trajectory optimization to reduce the likelihood of failure in case of unexpected connection conditions between flight phases.

Failure of the propulsion system is the most frequent cause of flight failure. From 1990 to 2015, 64 failures were reported worldwide due to problems with propulsion systems, accounting for 51% of total failures. Most launch failures are likely to be remedied by advanced GNC technology.

On December 5, 2010, three GLONASS satellites failed to enter into orbit. At IAF’2016, Russian researchers proposed adaptive optimal guidance as a solution to such problems[16], where autonomous online “end-toend” trajectory planning is the key technology.

In China, cases of mission failure, or excessive deviation from the target orbit have occurred recently because of the abnormal termination of thrust. For example, in the 22nd launch of a rocket, owing to the abnormal operation of the third-stage engine following the second ignition (insufficient thrust by gas clamping), the satellite entered the Earth atmosphere over Antarctica.

Here an example is useful. If the second-stage engine of a launcher shuts down before time (280 s),the thrust declines in advance compared with that in normal conditions. The launcher may thus not enter the target orbit (200-km circular orbit), and it may be impossible for it to form any orbit with the original guidance method, as shown in Figure 2(a). If a fault is detected and the guidance method is adjusted in time,the launcher can enter an elliptical orbit of 160 km ×180 km, as shown in Figure 2(b). The satellite can then enter into the target orbit through its own maneuvering, or can wait to be rescued.

Figure 2 Effect of rescue orbit during thrust reduction failure

Unlike traditional optimization problems, the rescue problem can be described as follows:

Goal: Approach as “close” as possible to the original target orbit.

Process constraints: Full use of propellants (completely exhausted), and attitude angle, angular rate, and height constraints:

Actually, autonomous guidance and precision landing has become a prerequisite for pinpoint landing on Mars[17]. The Mars Science Laboratory (2011) was the first Mars mission to attempt guided entry to safely deliver the rover to a touchdown ellipse of 25 km × 20 km[18]. Nowadays, however, requirements of landing accuracy have changed from tens of kilometers to hundreds of meters.

The powered landing problem can be described as follows:

Goal: minimal fuel for return

Eq. (13) above is closely coupled to the launcher.

The complexity of its constraints and solutions thus significantly increases.

3) For uncertainty in the models, parameters, and disturbances, the adaptive, robust control and ARDC methods would be replaced by AI technology, or inflight modeling and correction based on distributed sensing,to improve flight adaptability.

3.3 Solutions

Of the three scenarios introduced in Section 3.2, 1) and 2)are at Level 1 of the E3model and 3) is at Level 0. Some solutions are as follows:

1) (For Level 0) AI algorithms can be applied to solve these problems, such as:

Chebyshev neural network (CNN) for attitude tracking control: A CNN is used to estimate the total disturbance in the system, thus effectively reducing chattering in the sliding-mode system.

Data mining for optimization: Map reduction is applied to process large amounts of data to extract useful information. Subsequently, the parameter optimization of guidance and control is completed based on association among the acquired data.

Ant colony algorithm for optimization: When the fitness function is stable before and after an iteration, the local minimum is avoided by adjusting the parameters and enhancing randomness.

Probabilistic NN for fault diagnosis: NN and empirical-mode decomposition are used for fault diagnosis of electromechanical equipment.

Chaotic particle swarm optimization: A method based on constrained PSO and the Powell optimization algorithm is used to solve the optimal gliding trajectory.

Artificial immune and GAs for maneuvering: The problem-solving capability of GAs and the memory retention characteristic of the immune system are combined to make short-term decisions in a dynamic environment to meet short-term goals.

2) (For Level 0) Perception capability can be enhanced to solve originally uncertain problems.

The launcher has an elastic characteristic with the propellant swaying, and it is not adequate to measure the motion of only the center of mass. Some solutions have been introduced: (1) mathematical modeling,where computation complexity, mass approximation,and characteristic frequency errors are inevitable; (2)modal testing, where the cost is high, and the parameters are also challenging to check and the results cannot be directly used for real-time control.

A new approach, online modal identification through distributed sensing systems, can be considered.The systems may sense local vibration, overload, structural deformation, and direct force, and should be easy to deploy owing to their low cost, simple interconnection, and array distribution. Fiber Bragg Grating (FBG)is a suitable candidate for the systems, and can achieve intelligent control and assist in elastic control and instability prediction.

FBG-based control can be implemented as follows:

· A modal estimation algorithm is used to estimate the strain modal function and modal frequency based on multi-point strain measurement.

· A displacement modal function is estimated according to the strain modal function and the generalized coordinates of each mode are calculated.

· The displacement, acceleration, and angular rate for each point of the beam model are obtained.

· A control model is established and evolved in real time.The above procedure converts stress/temperature into 3D deformation. Modal/frequency identification,model/inertial measurement correction, and attitude deviation determination, are then processed in turn.The control law is then adjusted based on the above processing. By training offline, the processes from modal/frequency identification to attitude deviation determination can be replaced by the NN for reflexive control. Moreover, real-time modal identification can refine the control model and relax the requirement for the installation of rate gyros.

3) (For Level 1) The problem can be solved directly by computation, or transformed to avoid the obstacles that may be encountered in a direct solution.

The initial guess for the direct solution is discussed in the next section.

3.4 Online Planning-based Autonomous Flight

3.4.1 Trajectory planning and guidance

Guidance methods can be classified into the following four categories: reference tracking, that is, following a pre-defined track; in-flight reference generation and tracking, which means generating a real-time reference trajectory and following that track; in-flight control search, i.e., one-dimension search usually to numerically solve equations of motion; and in-flight optimal control, requiring numerical methods to satisfy some cost function.

Table 4 lists the differences between trajectory planning and guidance, where the former is usually carried out offline[19]:

Online trajectory planning, or autonomous guidance, seeks a balance to bridge the gap between trajectory optimization and guidance.

3.4.2 System configuration

The functional configuration of flight control based on online planning is shown in Figure 3. The capability assessment, target selection, constraints process and dynamic planning are landmark features.

Inflight modeling involves determining whether the flight is in a normal condition. If not, the failure mode should be identified and flight model updated. Capability assessment involves determining whether the target matches the given situation, especially when thrust drops. It is important to note that the original target may still be reached with a slight reduction in thrust.An analysis is then conducted to determine if the constraints need to be relaxed and how to relax them. Both strategies are intended to meet the requirements of real-time planning.

Table 4 Differences between trajectory optimization and guidance

Figure 3 Functional block of an intelligent flight system

If the constraints are simple, they can be met during planning; if not, the flight can be segmented, where tracking is preferable for segments with complex constraints, while online planning is adopted for simple constraints to clear all control errors accumulated in the preceding segments.

3.4.3 Capability assessment

The process of capability assessment is shown in Figure 4,and is based on analytical guidance, or iterative guidance. However, some simplifications are needed to determine rescue orbits:

1) The orbital plane parameters, i.e., the inclination of the orbit (i) and the longitude of the ascending node (Ω),are kept the same as that when failure occurred.

2) A circular orbit is selected for simplified calculation, i.e.,the eccentricity e = 0.

Figure 4 Ability assessment functional block diagram

3.4.4 Dynamic trajectory optimization

In Ref.[20], a solution framework was proposed for dynamic trajectory optimization. The effect of dynamic trajectory planning is shown in Figure 5. In each control cycle,flight trajectory was recalculated. The solid line in the figure shows the actual flight trajectory and the broken line the planned trajectory. It is clear that the flight did not track the planned trajectory very well, but in each control cycle, a new trajectory replaced the old one according to the given state. Therefore, although control deviations always existed,the errors no longer accumulated, and the final accuracy depended on the deviation following the newest trajectory planned in the last cycle, and the adjustment capability of the launcher.

This scheme relaxes the requirements of attitude control for trajectory tracking. However, the planning process should ensure real-time and convergent results, where the initial guess plays a vital role.

3.4.5 Initial value generation

The initial value can be determined as follows:

1) Prediction and correction with engineering experience or simulation analysis It is inefficient, and relies on experience to a significant degree.

2) Selecting a suboptimal solution of simplified, approximated, or transformed problems through one of the following:

- The suboptimal solution can be obtained by an analytic guidance law.

- The solutions of simplified models are set as the initial value of the problem with respect to the original models. The solution can be obtained by the multiple shooting method, the collocation method, and the least squares approach. Smooth deformation is then applied to solve the original problem[21]:

Figure 5 Illustration of online planning

Where F1(X) is the original model, F2(X) is the simplified model, and F(X,λ) is the combined model.

Whenλchanges from one to zero, F(X,λ) changes from the simplified to the original model.

- The solution of simplified constraints is set as the initial value of the problem with all constraints in place[22].

3) Using evolutionary and other heuristic methods

The initial guess can be generated quickly based on an intelligent algorithm trained on the ground, which is closely based on the ideas introduced in Section 1.2

It is reasonable to select a pre-designed trajectory as the first initial value in online planning. Due to the planning in each control cycle and the short interval (usually tens to hundreds of milliseconds)between control cycles, tracking deviation should be small even if attitude control cannot accurately follow the guidance command. Thus, the interpolation from the planned result of the previous cycle can be set as the initial value of the given cycle. If this approach can ensure convergence, or a trust region of the initial value is found where the locally convex or convergent requirements are met, the obstacles to real-time planning are eliminated. Figure 6 explains the trust region.

3.5 Application

The powered landing problem introduced in Eqs. (7) - (15)is discussed here. Two scenarios, without any constraints (blue)and with some constraints (red), were compared. From the last figure in Figure 7, it is clear that the thrust control exhibited the bang-bang characteristic. The engines only restarted once for retro-propulsion under no constraints. Corresponding to this, the engines restarted twice more to meet the constraints of dynamic pressure and thermal flow. From this external performance, the planning revealed a certain degree of intelligence and autonomy, i.e. computational intelligence.

Figure 6 Trust region of initial values

Figure 7 Planning result without/with constraints

4 CONCLUSION

Limited by various resources and the demand for risk aversion, intelligent and autonomous technology in space flight is not widespread, butis commonly used in such information perception tasks such as navigation, image processing and matching,health management, and fault diagnosis.

Using current computing power and sophisticated algorithms, numerical optimization in flight is nearly impossible for problems with complex constraints and dynamic models. However, according to the physical features and the targets of different flight stages, it is possible to apply computational guidance in flight by simplifying the models and constraints. The following steps are thus preferable for engineering applications:

1) offline design to optimize the computational framework;

2) improving optimization cores to enhance computational efficiency according to the specific problem; and

3) embedded algorithm optimization.

The requirements of autonomous trajectory planning and guidance have become more pressing over time, and will eventually become a prerequisite for space exploration. The levels of E3reflect current understanding of the intelligent and autonomous flight of launchers, and indicate some solutions, where computational intelligence and distributed sensing are expected to be the two main means to this end.