ALPSFORT - A Learning graPh-baSed framework FOr cybeR-physical sysTems
Advances in embedded systems and communication technologies, together with the availability of low-cost sensors, have paved the way to a pervasive presence of cyber-human applications in our everyday lives as well as in diversified segments of the market, e.g., those associated with the Internet-of-Things and Objects (IoT) and Cyber-Physical Systems (CPS). Such applications are mostly data-eager, in the sense that application decisions and the interaction with the physical world are strongly driven by the information content extracted from a (possibly very) large platform of sensors. This dependency on sensor data can be, however, a point of failure for the envisaged applications, since sensors and electronic apparatus are prone to faults and malfunctioning. This, in turn, negatively affects the information content carried by data that is used by the application to make decisions. At the same time, acquired data show to be time-variant in many real application scenarios, as the interaction between sensors and the environment and/or the environment itself evolve with time. The immediate consequence of time-variance is that, without any envisaged adaptation mechanisms, the application performances degrade and might become quickly unacceptable. Here, we aim at relaxing two major assumptions that largely impact on cyber-human applications and prevent their credibility whenever devices are either deployed in harsh –critical- environments or it is requested them to guarantee performances over time: 1) the stationarity/time-variance hypothesis of acquired data, and 2) the data correctness. By relaxing the former hypothesis we enter in the realm of the research known as “learning in nonstationary environments”; once we relax the latter one, we face the “fault diagnosis systems” research framework. This project aims at tackling in a comprehensive way both stationarity/time-variance and fault diagnosis aspects within a unique methodological framework. Our goal is to provide methods for designing intelligent embedded and CPS able to 1) detect the presence of changes in the system-environment interaction; 2) disambiguate among changes induced by time-variance, false positives, and faults affecting sensors; 3) identify the characteristics of the change and isolate it; 4) introduce reaction mechanisms to time-variance and mitigate the occurrence of faults at the sensor level. Here, we aim at providing a methodology able to automatically learn suitable/useful information needed by detection, identification, disambiguation, isolation and mitigation methods -and introduce corrective actions- directly from the data the application receives. This is known in the literature as “model-free” approach: no model is requested to describe the system/environment under investigation, no change/fault dictionary or signatures are expected, no a priori information about the nature of uncertainty or available sensors provided. All unknown entities needed to design CPS responsive to changes in stationarity/able to accommodate faults are learned directly from available data through computational intelligence and machine learning techniques. Networked embedded systems, CPSs and IoT natively support a graph-based representation. For instance, we can create graphs by inspecting technological, application or data-related properties of CPSs, e.g., the communication activity among units, the set of sensors used by the application to solve a specific task, the existing dependency in time and/or space among sensors. This unique characteristic permits to address the identified research problems with a completely novel set of tools relying on graph-based machine learning approaches, so as to take advantage of information associated with changes affecting the network topology, the existence of distributed information, and evolution of system behaviour. This approach is completely novel and represents the core of the project, whose outcomes will support the design of a new generation of intelligent distributed embedded systems and cyber-human applications.
- Zambon D., Grattarola D., Livi L., Alippi C. (2019) Autoregressive Models for Sequences of Graphs. IEEE. IEEE International Joint Conference on Neural Networks
- Grattarola D., Zambon D., Livi L., Alippi C. (2019) Change Detection in Graph Streams by Learning Graph Embeddings on Constant-Curvature Manifolds, IEEE Transactions on Neural Networks and Learning Systems:1-14
- Zambon D., Alippi C., Livi L. (2019) Change-Point Methods on a Sequence of Graphs, IEEE Transactions on Signal Processing:6327-6341
- Zambon D., Livi L., Alippi C. (2018) Anomaly and Change Detection in Graph Streams through Constant-Curvature Manifold Embeddings. 2018 International Joint Conference on Neural Networks. 2018 World Congress on Computational Intelligence. Rio de Janeiro, Brazil. 8-13 July 2018
- Zambon D., Alippi C., Livi L. (2018) Concept Drift and Anomaly Detection in Graph Streams, IEEE Transactions on Neural Networks and Learning Systems:5592-5605. ISSN 2162-2388
- Zambon D., Livi L., Alippi C. (2017) Detecting changes in sequences of attributed graphs. 2017 IEEE Symposium Series on Computational Intelligence (SSCI) Proceedings. IEEE. IEEE Symposium Series on Computational Intelligence. Honolulu, Hawaii, USA. 27 Nov.-1 Dec. 2017. ISBN 9781538627266