On coordination of cyber-physical systems

Abstract

The physical world will be saturated by networked devices with sensing and actuating capabilities. This trend will generate profound impacts on our way of monitoring and controlling the physical world, and meanwhile bring us towards a cyber-physical convergence era. Along this trend, a new type of systems, named Cyber-Physical Systems (CPS),is developed by taking into consideration the cyber and physical world interactions. CPS refer to the integrated systems with both cyber and physical aspects, and their applications have been envisioned to generate grand societal and economic impacts. In this dissertation, we study the requirements and fundamental problems in CPS design. More specifically, we address the issue of how devices in distributed CPS coordinate their activities in order to achieve different goals in CPS applications. Examples of the coordination goals include co-operation of devices for verifying and maintaining desired system properties, and adapting to external situations, etc. Traditional methods of dealing with coordination have become insufficient in the CPS scenarios due to CPS' special characteristics, such as tight coupling of cyber and physical components, and safety/time critical nature. Moreover, the internal and external environments of CPS are intrinsically associated with diverse uncertainties, such as variation of network conditions, unpredictable user interactions, etc. Those uncertainties further complicate the coordination. This dissertation actually has made original contributions in developing solutions to the above coordination problems. First, we study how CPS devices coordinate to verify the system's properties, like safety. Our study is conducted in a context of Medical Cyber-Physical Systems (MCPS), which are a typical CPS application in hospital and aim to coordinate networked medical devices for safer delivery of medical care to patients (i.e., a physical component). In medical domain, formal verification is a highly desirable procedure for eliminating design defects and increasing software quality. The combination of control, networking and patient in MCPS requires a joint modelling of the dynamics about the cyber and physical components, resulting in a hybrid system model. However, obtaining a tractable model for patient is difficult due to the uncertainties about patient's response to medical treatments (e.g., medication infusion). As a result, the overall verification is hindered. To solve this problem, our approach is to transform the traditional offine verification to an online version, in which MCPS devices co-ordinate at runtime to periodically generate a hybrid system model describing the system's behavior in the future short-term. Then, formal verification, specifically model checking, can be conducted over the online model. The rational under our approach is the fact that patient's physiological dynamics turns to be predictable in the near future and thus can be readily modelled. Furthermore, we propose to build the online verification procedure as a real-time task for continuous fault prediction and risk prevention. Once the online verification results indicate a violation of the desired properties, the devices coordinate to enter a fall-back plan, thus circumventing the potential risks. The feasibility and effectiveness of our approach have been validated in a concrete medical case, named Airway-laser Surgery. Second, we study how CPS devices coordinate to maintain a crucial system property, i.e., safety. The definition of safety varies in different application contexts. We restrict our study to MCPS where safety refers to freedom of medical accidents. To maintain this property, medical devices must communicate with each other, and act appropriately and promptly. However, one challenge that hinders the fulfillment of the safety property is network uncertainties, such as message loss and varied transmission delay. This problem arises especially when wireless technologies are adopted for inter-connecting medical devices. Existing coordination mechanisms in MCPS for dealing with network uncertainties are time-triggered and consequently can not give prompt response to surgeon's requests. The mechanism we proposed is event-triggered, and thus coordination can be initiated whenever needed. Besides, in order to tolerate network uncertainties, we introduce a central entity, called supervisor, to oversee the states of patient and each medical device, and then coordinate the current and future operations for each medical device. We designed the interaction mechanism for the supervisor and medical devices, as well as the coordination algorithm for the supervisor. In the algorithm, we take into account the special requirements from medical domain, and prove that the planned operations are always safe to be adopted by medical devices when network problems occur. Trace-driven simulation has demonstrated the advantages of our scheme in many aspects, such as clinical efficiency and response delay, compared to the existing approaches. Third, we investigate the problem of coordinating CPS devices to enable adaptation to external situations, like user status and environmental changes. Increasing the system's awareness to the external situations has many benefits, e.g., createing autonomous and intelligent CPS applications, and better serving users. Again, we choose MCPS as the investigation scenario, and propose to make MCPS particularly adaptive to user's status, because the safety and quality of medical care highly depend on user{174}s participation and operation. Specifically, we treat user errors, emerging when users are operating medical devices, as contexts to MCPS, and build a context infrastructure to infer the user errors. With recognition of these errors, devices in MCPS then can coordinate appropriately to prevent the potential risks of those errors. However, it has been well acknowledged that the contexts we obtain are subject to uncertainty. Existing work to handle context uncertainty mainly focuses on modeling, inference, and mitigation of uncertainty. The adverse consequences of uncertainty, especially on patient safety, are neglected. Our solution to addressing this issue is two-fold. First, it is necessary to improve the quality of the detected contexts by appropriate approaches. Second, we modify the "context-action" adaptation style in the traditional context-aware systems to a "context-assessment-action" style, where the "assessment" step evaluates the safety of each context-triggered action based on solid medical knowledge. An action is allowed to be executed only if its execution is ascertained to be safe. Following this principle, we have developed a context-aware MCPS for a particular medial case, named Patient-Controlled Analgesia (PCA). Evaluation results about the system have shown that our system can prevent 95% more adverse events compared to the state-of-the-art context-insensitive PCA systems.