The rapid development of computing technology and automation widens the scope of the task delegated to cyber-physical systems (CPS) such as smart grids or smart buildings. Explainability, i.e., the ability to provide explanations about system states or behaviors becomes one of the requirements for future cyber-physical systems as more complex computer-made decisions a ect our daily lives. The work on the explainability in CPS is scarce despite recent attention on the explainability of algorithms in arti cial intelligence. This doctorate research aims to comprehensively understand the scope of explainability in CPS, identify the critical components of an explainable CPS, and methods and metrics to evaluate them. Speci cally, our main research question is how and to what extent Knowledge Graphs can be applied in enabling the explainability of CPS. Using the design science approach, we attempt to answer these questions in a set of iterations, starting with a simulationbased approach and constructing a baseline system followed by more focused studies and more realistic settings using data from real-world CPS. The selected application domain in this work is industrial energy systems such as smart grids and smart buildings. The expected outcome of this work is a theoretical foundation and methods for developing an explainable CPS applicable in various domains.
Recently there has been concern that future CPS which span both the realm
of physical and cyber-worlds are challenged to explain their behavior to users,
engineers, and other stakeholders [
Explainability is the ability of a (software) system to provide explanations
about its states or behaviors in terms of a set of facts [
Exploring various risk-related scenarios is undesirable to be conducted in the real system, that is in vivo. Having a in vitro platform and reusable framework allows for a more rapid development process and avoiding the unnecessary cost of trial and error when developing an explainable CPS.
As an illustrative example in the energy domain, smart electricity grids evolve from static to dynamically changing networks of large numbers of devices, e.g., photo-voltaic units (PV), electric vehicle charging stations (EVCS). The slow charging of an EVCS is an event that requires an explanation for several stakeholders including the EVCS owner, customer service representatives, eld engineers, and grid planners. An explanation could be that overcast weather leads to lower than usual energy production through PVs in the region, this leads to a lack of supply in the grid segment and to a control intervention to reduce charging power by the grid operator. From the consumer's perspective, the change in energy consumption should be seen as independent of how it is produced. The energy production is then expected to run uninterrupted and provide su cient supply even when there is an increase in consumption. A swift response is desired if an unwanted event such as failure to ful ll expected service or blackout is unavoidable since the loss caused by the fault would be a function of time. On the one hand, the technical operation employees expect detailed explanations in order to be able to decide the next course of action to remedy a potential fault/anomaly. On the other hand, the possibly larger population of a ected consumers, an ideal explanation would be more succinct and related to their context, such as the service contract.
In order to generate perspicuous explanations for the intended stakeholders, the explanatory system needs to integrate information from various sources such as the structure of the system, the relationship between elements of the system, and the history of the system's state. Additionally, understanding the recipient of an explanation is also essential to be tailored to be easy to understand. 2
Arti cial intelligence, mainly the area of expert and knowledge-based systems,
extensively studies the task of providing explanation based on formalized logical
reasoning [
The interpretation of explainability from the perspective of industrial
systems is even more pragmatic. Related topics such as anomaly detection and
subsequent root-cause analysis are essential topics in industrial (cyber-physical)
systems and are currently achieved with methods such as FMEA (Failure Mode
and E ect Analysis) [
CPS, particularly the smart grid, is relatively new and evolving, and it
combines di erent disciplines such as physics, statistics, and socio-economics. Studies
of explainability in CPS are scarce, especially for speci c topics such as the
approaches based on the knowledge graph. One of the closest approaches is fault
diagnosis systems in a smart building that combines a physical process model
and data-driven approach[
In summary, explainability encompasses, on the one end, a human who needs an explanation and, on the other end, causality knowledge that is not known explicitly from the system's description. Existing literature addressed these issues only partially, and the focus of di erent communities is diverse. Only by collecting various puzzle pieces can we see the big picture and establish a solid foundation of explainable CPS. 3
The literature suggests that there is a limited understanding of what constitutes an explainable cyber-physical systems. One line of research focuses on only user aspect of explanation while other line struggles with ad-hoc or partial solutions. Therefore, the main question for this research is: RQ0 What are the main theoretical, methodological, and engineering foundations that enable e ective and e cient implementation of explainable CPS?
This question is the starting point to the more speci c questions: What are the core components needed to achieve explainability? To what extent knowledge graph can help build an explainable CPS? What are the requirements for making an explainable system applicable to various domains?
The literature also indicates that causality knowledge and user-oriented
explanation generation algorithms are the critical aspect of an explainable CPS.
Thus, this research also aims to address the following questions:
RQ1 What is the e ective semantic representation to integrate di erent
representations of causality knowledge? A di erent source of causality knowledge may
have a di erent meaning of weight of a causal relationship. Some may involve the
coe cient of a di erential equation, and some others might refer to probabilistic
quantities to refer to subjective belief or derived quality metrics.
RQ2 How to acquire causality knowledge e ciently from data and domain
experts? How can we support domain experts to express causal relationships based
on domain knowledge and data? One approach to express causality captured
from domain experts used in [
The questions above correspond to the core functionalities of explanation generation: causality knowledge acquisition and exploitation. Other aspects of the interface to the end-users, such as visualization or generation of explanation in natural language, will not be addressed in this research. These aspects will become more apparent when the core and realistic use case scenarios have been developed. 4
This study adopts a classical Design Science methodology [
This study has been conducted for a year. The following 2-3 years will be focused on answering each research question. Speci cally, the second year will be allocated for addressing the representation and acquisition (RQ1) of causality knowledge (RQ2). The analytics (RQ3) and presentation (RQ4) aspect of the explanation will be conducted in the third year.
Preliminary results The current state of the PhD has resulted in an
understanding of explainability and explainable CPS. In the rst iteration, the work is
oriented towards understanding the explainability in energy systems.
Developing plausible and feasible scenarios and data acquisition drives the focus on using
a simulation platform (i.e. BIFROST [
The following iteration builds upon the previous idea with more concrete
artifacts. One of the results is the architecture shown in Figure 1. The
realization of this architecture was then implemented as a prototype application for
demonstrating that the explanations from the scenario can be derived based on
simulated data and captured knowledge. The solution design and
implementation result was then published in the energy community proposing the solution
based on semantic web technologies [
Figure 2 displays the prototype of an explainable CPS build as a part of the BIFROST smart grid simulation engine.
Behind the user-facing interface is the engine that integrates data coming from the simulation into knowledge graphs in the triple store, deducing causal relations, detecting events, and deriving explanation for the detected events. The explanation is then displayed as shown on the right side of the gure. From this rst iteration, we better understand what aspects are needed to build an explainable CPS. 5
The goal of an explainable CPS is to provide explanations of events for a variety of scenarios. Close collaboration with domain experts is necessary To achieve plausible scenarios that can be used as a basis for further evaluation. The developed scenarios are then implemented in a simulation to generate data. Furthermore, actual measurements will be used to ensure the validity of the simulation data. User studies and empirical analysis are performed to evaluate the research questions. The following describes the outcome and outputs for each research question.
RQ1 The outcome for RQ1 is the incorporation of di erent causality representations to generate an explanation. A vocabulary for di erent meanings (e.g., 1 https://pebbie.org/expcps/ relationship weight) of causality will be designed to augment the basic model of causality. Additionally, an algorithm to fuse these di erent semantics of causality will be developed as part of the explanation generation algorithm. RQ2 Answering RQ2 will involve user studies and implementation of algorithms to derive causality knowledge from simulated data. Method to acquire causality knowledge will be developed and evaluated using user-study and empirical analysis.
RQ3 A set of metrics and ranking algorithms will be developed, and an
explanation generation task will be executed based on a prepared scenario. A group
of domain experts will be asked to manually create explanations as the gold
standard for measuring the algorithm's performance. The evaluation of the
algorithm will use metrics such as MRR or Precision@10 as a base and modi ed
to accommodate comparing the graph structure of the explanation.
RQ4 A qualitative study will be conducted to acquire key characteristics of
explanation target or user pro les. A set of user-pro les will be de ned, and
the explanation generation task will be executed using the scenarios. Another
user study will assess whether the customized generated explanation is relevant
to the intended explanation recipients. To this end, System Causability Scale
(SCS) [
The previous section described an end-to-end prototype of how an explainable CPS should work after the rst year of the doctoral study. This initial work helps to identify issues formulated in the research questions for the next iteration. The collection of artifacts from investigating each research question forms a solution framework to build an explainable CPS.
Some topics possibly related to explainable CPS are intentionally not addressed considering the limited scope of this research. e.g., such as scalable storage techniques to handle large-scale CPS data. Other topics depend on the mentioned research questions, such as the study of speci c presentation forms of explanation (e.g., visualization) or exploratory search systems to explore the explanation hypothesis and history of events. Further research on these topics will enrich the framework for building an explainable CPS and enables explainability in various systems. 7
This work is funded through a grant project titled "Power System Cogni cation" (PoSyCo) by the Austrian Research Promotion Agency (FFG) with Siemens AG. I am thankful for PhD supervisory team Marta Sabou, Fajar Juang Ekaputra, and Tomasz Miksa for the professional advises and personal supports.