INTELLLIGENT SYSTEMS ROADMAP
Topic Area: Adaptive and Non-Deterministic Systems
Christine Belcastro, NASA LARC
Nhan Nguyen, NASA ARC
1. Roles and Capabilities
As demands on aerospace accessibility increase and become more complex, intelligent systems technologies can play many important roles to improve operational efficiency, mission performance, and safety for current and future aerospace systems and operations. Future intelligent systems technologies can provide increased adaptive and autonomous capabilities at all levels, as illustrated in Figure 1. At the lowest level of autonomy, adaptation through closed-loop control and prognostics enables aerospace systems to be more resilient by automatically adjusting system operations to cope with unanticipated changes in system performance and operating environment. At mid-level of autonomy, planning and scheduling provide capabilities to perform automatic task allocations and contingency management to reduce human operator workloads and improve situational awareness and mission planning. At high levels of autonomy, automated reasoning and decision support systems provide higher degrees of intelligence to enable aerospace systems to achieve autonomous operations without direct human supervision in the loop.
Figure 1. Levels of Autonomy Integration with Human Operators
Adaptive systems are an important enabling feature common to all these levels of autonomy. Adaptive systems can learn and optimize system behaviors to improve system performance and safety. Adaptive systems can also enable efficient, intelligent use of resources in aerospace processes. Furthermore, adaptive systems can predict and estimate aerospace system’s long-term and short-term behaviors via strategic learning and tactical adaptation. As a result, safety improvements through resiliency can be achieved by adaptive systems, which can automatically perform detection and mitigation of uncertain, unanticipated, and hazardous conditions, thereby enabling real-time safety assurance.
As aerospace systems become increasingly more complex, uncertainty can degrade system performance and operation. Uncertainty will always exist in all aerospace systems, no matter how small, due to imperfect knowledge of system behaviors, which are usually modeled mathematically or empirically. Uncertainty can be managed but cannot be eliminated. Typical risk management of uncertainty in aerospace systems requires improved system knowledge by better system modeling which can be very expensive, built-in safety margins and operational restrictions which can sometimes adversely impact performance if safety margins are unnecessarily large or operational restrictions are not well established.
Adaptive systems can better manage uncertainty in aerospace systems if they are properly designed. Uncertainty is managed by adaptation, which adjusts system behaviors to changing environment through learning and adopting new behaviors to cope with changes. Adaptive systems provide learning mechanisms to internally adjust system performance and operation to achieve desired system behaviors while suppressing undesired responses, and to seek optimal system behaviors over long time horizon.
Adaptive systems achieve adaptation through short-term tactical adaptation and long-term strategic learning and self-optimization. Tactical adaptation usually involves the need to adjust system behaviors to cope with rapid changes in operating environments that could cause safety concerns. Model-reference adaptive control is an example of a tactical adaptation strategy that has many potential promises in future aerospace systems. Adaptive flight control has many applications in air and space vehicle systems.
Strategic learning and self-optimization are learning mechanisms of adaptive systems that can take place over a longer time horizon. This adaptation mechanism usually addresses the need to adjust system behaviors to optimize system performance in the presence of uncertainty. Examples of strategic learning are reinforcement learning and extremum seeking self-optimization in aerospace systems. Air and space vehicles can leverage extremum seeking self- optimization to adjust their flight trajectories and performance characteristics to achieve energy savings.
Adaptive systems can provide many useful applications in aerospace systems. Adaptive flight control for safety resiliency to maintain stability of aircraft with structural and/or actuator failures has been well studied. Real-time drag optimization of future transport aircraft is an example of extremum seeking self-optimization that can potentially improve fuel efficiency of aircraft. Adaptive traction control of surface mobility planetary rovers can be applied to improve vehicle traction on different types of surface. Adaptive planning and scheduling can play a role in air traffic management to perform weather routing or traffic congestion planning of aircraft in the National Air Space.
Machine learning techniques are commonly used in many adaptive systems. These techniques sometimes employ neural networks to model complex system behaviors. The use of multi-layer neural networks can result in non-determinism due to random weight initialization. Non-deterministic behaviors of these adaptive systems can cause many issues for safety assurance and verification and validation. Neural networks are not the only source of non-determinism. Stochastic processes such as atmospheric turbulence, process noise, and reasoning processes such as due to diagnostics/prognostics can also be sources of non-determinism.
Adaptive systems offer potential promising technologies for future aerospace systems. Nonetheless, many technical challenges exist that prevent potential benefits of adaptive systems from being fully realized. These technical challenges present technology barriers that must be addressed in order to enable intelligent systems technologies in future aerospace systems.
2. Technical Challenges and Technology Barriers
Technical challenges and technology barriers must be defined at all levels of autonomy integration with human operators for ensuring safety, operational efficiency, and improved performance. Figure 2 illustrates this concept for aircraft systems, with levels of autonomy integration associated with a potential timeframe for implementation. At the lowest level of integration to support existing baseline systems and human operators, adaptive and reasoning systems can improve safety through resilience under uncertain, unexpected, and hazardous conditions by providing improved situation awareness, guidance, and temporary interventions under emergency conditions. At a mid-level of integration, semi-autonomous systems can enable synergistic dynamic teaming between human operators and intelligent systems to improve safety and operational efficiency. At the highest level of integration, fully autonomous systems can ensure safety and optimize operational efficiency and performance, while keeping a (possibly remote) human operator informed of current status and future potential risks.
Figure 2. Illustration of Adaptive Systems Multi-Level Role for Aircraft
Technical challenges and technology impediments are summarized below for improving safety, operational efficiency, and performance at all integration levels.
2.1. Improved Safety
Key technical challenges and technology impediments for achieving resilience of safety-critical aerospace systems at all integration levels under uncertain, unexpected, and hazardous conditions are summarized below.
2.1.1. Technical Challenges:
· Development and Validation of Resilient Systems Technologies for Multiple Hazards
- Reliable Contingency Management (Control & Routing) for Unexpected Events
- Accurate & Fast Situation Assessment, Impacts Prediction, and Prioritization
- Fast Decision-Making and Ensured Appropriate Response
- Real-Time Sensor & Information Integrity Assurance
· Development and Validation of Variable Autonomy Systems that Enable Effective Teaming between Human Operators and Automation
− Standard, Effective, and Robust Multiple Modality Interface System
− Common Real-Time Situation Understanding between Human and Automation (Including standard taxonomies and lexicon of terms)
− Real-Time Dynamic Effective Task Allocation and Decision Monitoring
· Development and Validation of Ultra-Reliable Safety-Assured Autonomy Technologies
− Real-Time Safety Assurance
− Universal Metrics and Requirements for Ultra-Reliable Safety-Assured Autonomy
− Hierarchical Integration and Compositional Analysis between Control and Planning Functions
2.1.2. Technology Impediments
· Certification of Safety-Assured Autonomy Systems for Reliable Operation under Uncertain, Unexpected, and Hazardous Conditions
· Integration into Existing Flight Deck Equipment and Operational System (e.g., Air Traffic Management (ATM) System)
· Public and Policy Perceptions associated with a Lack of Trust in Autonomy Technologies
· Cyber Security (both a Challenge and an Impediment)
· Lack of Alignment and Integration between Control, Artificial Intelligence (AI), and Software Validation and Verification (V&V) Communities
2.2. Improved Operational Efficiency
Key technical challenges and technology impediments for improving operational efficiency for semi-autonomous and fully autonomous aerospace systems are summarized below.
2.2.1. Technical Challenges
· Real-Time Decision Support and Mission Planning for Single Pilot Operations
· Pilot Monitoring and Decision-Making for Impaired Pilot (or Human Operator)
2.2.2. Technology Impediments
· Certification of Adaptive Systems and Automatic Takeover of Control Authority under Pilot Impairment
· Integration into Existing Flight Deck Equipment and ATM System
· Public and Policy Perceptions associated with a Lack of Trust in Autonomy Technologies
· Cyber Security (both a Challenge and an Impediment)
· Lack of Alignment and Integration between Control, Artificial Intelligence (AI), and Software Validation and Verification (V&V) Communities
2.3. Improved Performance
Key technical challenges and technology impediments for improving performance for semi-autonomous and fully autonomous aerospace systems are summarized below.
2.3.1. Technical Challenges
· Real-Time Drag, Aerodynamic, Aeroservoelastic, and Aeropropulsive Optimization
· Real-Time Optimization, Convergence, and Computational Intensity
2.3.2. Technology Impediments
· Certification of Adaptive Systems for Unstable Aerospace Vehicles
· Integration into Existing Flight Deck Equipment and ATM System
· Public and Policy Perceptions associated with a Lack of Trust in Autonomy Technologies
· Cyber Security (both a Challenge and an Impediment)
· Lack of Alignment and Integration between Control, Artificial Intelligence (AI), and Software Validation and Verification (V&V) Communities
These technical challenges and technology impediments define research needs that must be addressed.
3. Research Needs to Accomplish Technical Challenges and Overcome Technology Barriers
Despite many recent advances, adaptive systems remain at a technology readiness level of 5. The furthest advancement of this technology has been flight testing on piloted research aircraft and subscale research aircraft under high-risk conditions, but no production safety-critical aerospace systems have yet employed adaptive systems. The existing approach to adaptive control synthesis generally lacks the ability to deal with integrated effects of many different (multidisciplinary) flight physics. In the presence of vehicle hazards such as damage or failures, flight vehicles can exhibit numerous coupled effects that impose a considerable degree of uncertainty on the vehicle performance and safety. To adequately deal with these coupled effects, an integrated approach in adaptive systems research should be taken that will require developing new fundamental multidisciplinary methods in adaptive control and modeling. These multidisciplinary methods in adaptive control research would develop a fundamental understanding of complex system interactions that manifest themselves in system uncertainty. With an improved understanding of the system uncertainty, effective adaptive systems could be developed to improve performance while ensuring robustness in the presence of uncertainty.
Another future research goal is to develop simplified adaptive systems that reduce the introduction of non-determinism. Despite the potential benefits of neural network applications in adaptive systems, real-world experiences through flight testing seem to suggest that simplified adaptive systems without neural networks may perform better in practice than those with neural networks. Simplified adaptive systems may have other advantages in that they may be easier to be verified and validated, and there are some existing adaptive control methods that can be applied to assess stability margins.
Applications of real-time self-optimization systems are still very limited, but the potential benefits of these systems can be enormous. Aircraft with self-optimization can potentially achieve significant fuel savings when equipped with suitable control systems that enable self-optimization. Research in methods of real-time extremum seeking self-optimization is needed to advance the technology to a level where it can consistently demonstrate reliability and effectiveness.
Research is needed in the development and validation of resilient control and mission management systems that enable real-time detection, identification, mitigation, recovery, and mission planning under multiple hazards. These hazards include vehicle impairment and system malfunction / failure, external and environmental disturbances, human operator errors, the sudden and unexpected appearance of fixed and moving obstacles, safety risks imposed by security threats, and combinations of these hazards. Resilient control and mission management functions include adverse condition sensing, detection, and impacts assessment, dynamic envelope estimation and protection, resilient control under off-nominal and hazardous conditions, upset detection and recovery, automatic obstacle detection and collision avoidance, and mission re-planning and emergency landing planning. Metrics and realistic current and future hazards-based scenarios are also needed for resilience testing of these systems, including a means of generating these hazards with an element of surprise during testing with human operators.
Research into variable autonomy systems is needed to facilitate dynamic effective teaming between the automation and human operators. Specific areas of research include real-time dynamic function allocation and interface systems, resilient guidance and autonomous control systems for loss of control prevention and recovery, as well as diagnostic and prognostic systems that enable information fusion, guidance and decision support under complex off-nominal and hazardous conditions.
Research is also needed for the development and validation of supervisory and management systems that enable real-time safety assurance of resilient, semi-autonomous, and fully autonomous systems operating under uncertain, unexpected, and hazardous conditions. These systems would monitor current vehicle and environmental conditions, all information provided by and actions taken (or not taken) by human operators and integrated intelligent systems (including vehicle health management, resilient control and mission planning), and assess the current and future safety state and associated safety risks in terms of multiple safety factors (including vehicle health and airworthiness, remaining margin prior to entering a loss of control condition, and time remaining for recovery). This capability would require deterministic and stochastic reasoning processes as well as an ability to reliably and temporarily intervene if necessary over both human and intelligent automation systems, while providing situational awareness and guidance to both.
Verification and validation (V&V) research is viewed as a key research to enable adaptive systems to be operational in future flight vehicles. V&V processes are designed to ensure that adaptive systems function as intended and the consequences of all possible outcomes of the adaptive control are verified to be acceptable. Currently, software V&V research is being conducted but is not aimed at adaptive systems or the functional validation of these systems at the algorithm level. To effectively develop verifiable adaptive systems, the roles of adaptive systems theory as well as aerospace system modeling should be tightly integrated with V&V methods. Otherwise, the V&V research could become stove-piped, thereby resulting in challenges in implementation.