Chapter I: Introduction

Origin and History of the Problem

Since the invention of the first automated systems during the Industrial Revolution, advances in sensor technology [14, 15, 16, 17, 18], computing power [19, 20, 21, 22], and communications [23, 24, 25, 26, 27] have dramatically increased the opportunities to incorporate devices and equipment into ever more complex systems. These innovations have helped entire industries to evolve and new business models to emerge in domains such as energy [28, 29], transportation [30, 31, 32], health [33, 34, 35], security [36, 37, 38], and entertainment [39, 40]. Many of these solutions are what today make up power systems, water supplies, and digital communication networks, among others. However, automation has also played an important role in undesired events and fatal accidents. Such is the case of the Air France Flight 447 that intended to transport 216 passengers and its 12-person crew from Rio de Janeiro to Paris on June 1, 2009 [41]. In this accident, the Airbus A330-200 crashed into the Atlantic Ocean after the aircraft entered an aerodynamic stall caused by erroneous readings from the airspeed sensors. These readings triggered a sequence of events that resulted in the disengagement of the autopilot, which left the control of the aircraft to the crew. At that moment, the crew had insufficient awareness of the situation to react properly to the flight condition.

In the commissioning and life-cycle of heterogeneous dynamic systems (HDS), composed of humans, physical systems, and computer agents, accidents are not just a consequence of the failure of a particular device or the collapse of an automation system. Humans, as well, play a role in most of these accidents, with human error 1 being responsible for 60%-90%, of the accidents reported in domains such as process control, aviation, and health care [3, 42, 43]. In particular, a review of military aviation mishaps [44] and a study of accidents in major air carriers found that 88% were caused by human error resulting from a lack of situation awareness [45]. As this example highlights, beyond purely technology-based problems, a number of issues exist in the integration of humans and automation technology [2, 8], with a lack of situation awareness being an important cause of human errors.

This work does not intend to overcome the causes of the Airbus accident in Air France Flight 447; in general, issues with human-automation systems are complex and dependent on their specific application domain. However, this research aims to make a contribution by informing tools of decision making with methods of computational intelligence and principles in cognitive engineering for the safe operation and automation of HDS. The theoretical objective is to provide a methodology for the integration of human-automation systems that generates a representation of the situation knowledge base (SKB) of human-experts in order to assist other operators, who may not necessarily be experts, in the operation and supervision of such systems. These e orts may result in user interfaces that provide operators with additional information in support of real-time decision making.

Life Support Systems and Their Challenges

One HDS of particular interest is regenerative life support systems (RLSS). These systems grow in importance with the development of long-duration human space exploration systems. The capability of habitation systems to regenerate life support consumables, such as oxygen, is one of the challenges of long-duration human space flight [46]. Such capability would reduce the frequency of resupply missions and presumably also reduce their operation cost. An example of such e orts is the 2 commissioning of the Water Recovery System (WRS) in the U.S. segment of the International Space Station (ISS), which recycles waste liquids, including urine, back into potable water [47, 11, 48]. Indeed, on April 10th 2010, at the Kennedy Space Center, President Barack Obama pronounced his \Remarks on Space Exploration in the 21st Century," and in his speech he included life support systems as a key technology to be developed for future long-duration space ight missions:

"And we will extend the life of the International Space Station likely by more than ve years, while actually using it for its intended purpose: conducting advanced research that can help improve the daily lives of people here on Earth, as well as testing and improving upon our capabilities in space. This includes technologies like more efficient life support systems that will help reduce the cost of future missions."

RLSS combine physico-chemical and biological processes to transform metabolic byproducts back into consumables. Their purpose is to increase the autonomy of the space habitat and to maintain an acceptable quality of life for its living organisms by properly reusing byproducts and regenerating consumables. But these processes require energy and time to transform byproducts and nutrients into consumables and edible products. Consequently, their maintenance imposes considerable workload to operators. In particular, their monitoring and automation poses a challenge: material loop closure may promote unintended interactions between chemical species within the habitat, potentially leading to the accumulation of unexpected chemical compounds that may affect individual life-support processes or even crew health. An example of such unintended chemical interactions is found in the 2010 WRS anomaly, caused by the accumulation of dimethylsilanediol (DMSD) in the Water Processing Assembly (WPA) of ISS [11]. This anomaly served as a good example of the disconnect that in some cases may be apparent between humans and automation, thus 3 becoming a potential cause for conflicts that may be addressed if considered as an issue between humans and automation. We can address this problem of disconnect by detecting such anomalies early enough and using measurements to present sensor information that improves the observability of these errors in monitoring and automation systems to the human operator. By addressing this problem, we seek to minimize human errors while increasing system reliability.

Situation-Oriented Automation of RLSS

The availability of new chemical and biological sensors, together with computational resources, enables the development of automation systems aimed to alleviate human workload, avoid human error, and increase overall reliability of RLSS. Beyond methods in robust [49] and adaptive control [50, 51], paradigms in switched control [52, 53, 54, 55] offer advantages for the management of the uncertainty caused by material loop closure. Switched control introduces attributes of flexibility and modularity to the control system [55]. These attributes may be used to allow for different control actions depending on the operational condition of the physical system and its situation in a given context. The situation can be understood as the subjective state of a system in relation to its context, which can be in itself defined by the environment and active goals.

The combination of abundant sensor information creates a sensing space in which these situations may be defined, which when detected may be used to influence the operation of the system toward a specific goal. Furthermore, the ability to detect known situations in the sensing space may also open opportunities to detect unknown situations, in which case the automation system may alert and request a human expert to perform observations, collect data, or intervene. In this case, the human expert would contribute his/her knowledge to the situation, i.e. his/her situation awareness, defined as the \perception of the elements in [his/her] environment within a volume of time and space, the comprehension of their meaning, and the projection of their status into the near future" [56, 57]. This work proposes a situation-oriented approach to the switched control paradigm that performs a quantization of the sensing space to allow the automation system to actively probe for information [54]. It employs a granular decomposition [58, 59]of the measured or estimated variables, with each granule defining a situation in which a specific control objective governs the RLSS. The granular decomposition is made consistent with the SKB of a human-expert through an adaptation process that enables the automation system to \learn" the human perception of each situation, or the concept of a particular situation, from sensor readings, measurements, and inputs from an human expert. An advantage of the consistency between the granular decomposition and the human perceptions of situations is the capacity it provides to the system for the development of coordination strategies between the human and automation, i.e. the human-automation system. In particular, this dissertation makes use of a reactive agent architecture based on fuzzy associative memories (FAM), or FAM-based agents, composed of n-dimensional non-interactive fuzzy sets [60, 61]. The methodology aggregates sensor information and a set of human-expert situation assessments to obtain a parametric representation of their SKB. Data sets collected by human experts are a raw and uncompressed representation of their knowledge about the system. These data sets can be obtained from individual human experts or crowdsourced to a group of them.

Summary

The overall goal of this work aims to contribute to the methodological development of situation-oriented and user-centered design approaches for the integration of humans and automated RLSS. Despite the fact that many of the processes involved in RLSS 5 can be automated, these systems still require humans-in-the-loop for monitoring and intervention. Furthermore, because not all their functions may be fully operationalized, RLSS needs the ability and involvement of the human operators. The approach used in this work consists of collecting data from experts and implementing control policies suitable for human-system interaction. As such, this work makes five main contributions:

  1. The design, modeling, and simulation of a ground-based platform for research in RLSS (Chapter 3). This ground-based platform provides a testbed for validation.
  2. The development of a a granular approach to the automation of regenerative life support systems (Chapter 4) to enable the management of control policies based on situation.
  3. The development of an aggregation algorithm to obtain situation knowledge bases from human experts (Chapter 4) that is used to automate system processes.
  4. The characterization and validation of the aggregation algorithm employing data sets from simulation and human participants (Chapters 5 and 6). It provides observations and recommendations on data set requirements based on ideal conditions and o ers validation making use of data sets produced by human participants.
  5. The exploration of data-set combination techniques based on granular computing to obtain crowdsourced situation knowledge bases (Chapter 7). It demonstrates the advantage of employing techniques that operate on the situation knowledge base of individuals after these have been aggregated with the algorithm instead of combining them as raw data sets.