Research Activity

Critical infrastructures, distributed sensor/actuators systems, social networks and public protection applications are examples of systems characterized by high complexity and production of big and uncertain data. As such, solutions designed to address specific applications require sophisticated mechanisms to grant data handling and process understanding, robustness and resilience abilities, capacity to detect changes in nonstationary and adapt to concept drift, self-awareness to diagnose a fault and self-healing mechanisms to repair it as well as support remote controllability and reprogrammability of the solution.

Moreover, we neither can further accept strong hypotheses that make the mathematics amenable at the cost of loss in effectiveness and applicability nor decouple the design of an intelligent cyberphysical system from reality and its implementation and deployment.
A multidisciplinary approach is needed at the system/system of systems level requiring the introduction of intelligence and adaptation abilities directly in the design phase of the solution. Here, machine learning and computational intelligence are precious tools, combined with traditional techniques, to address and solve the above aspects yielding credible solutions, transferable to industry. 

Machine learning techniques have been constituting the leitmotiv of the study and lead to the design of intelligent systems and smart solutions to not-trivial problems and real-world applications.
My research goes in the above outlined direction by coupling machine learning mechanisms with adaptive processing systems so as to deliver a new generation of intelligent systems characterised by self-healing, decision making and adaptation abilities.
Current ongoing research addresses issues related to machine learning in non-stationary and evolving environments and intelligent embedded systems; research is carried out both at academic and industrial level.


The three main research addresses are the following:


Adaptive Intelligent Systems

The current research focuses on the theory, implementation and applications of learning machines  embedding adaptation mechanisms. Results and developed methodologies shed light on the structural and functional properties underlying such complex systems as well as address the performance/constraints trade-off when an application is envisioned.

Efforts are devoted to nonparametric change detection tests designed to detect concept drift affecting  datastreams. The methods detect changes in sequential data, not necessarily structured, by differentiating changes associated with nonstationarity, faults or model bias within a cognitive framework. Adaptive mechanisms allowing the system to react just in time to the perceived changes, i.e., exactly when it is needed, are also object of the research with a specific focus on classification systems and coding-decoding recurrent machines. Results have been applied to several applications, e.g., explosive and drug detections from X rays imagery, molecular explosive detections, photovoltaic maximum-power-point-tracker energy harvesting, laser welding and cutting, quality analysis applications.


Intelligent Embedded Systems

The research addresses methodological and application-related aspects of Intelligent cyberphysical embedded systems, i.e., embedded systems with sensors and actuators, executing computational intelligence algorithms to deal with uncertainty and learn from incoming sensor data. The class of embedded systems known in the literature as Wireless Sensor Networks, Internet of Things, passive RFId-based and hybrid systems are object of the study. More specifically, aspects related to energy harvesting and storage, energy management (energy-aware routing protocols, unit management, adaptive sampling, dynamic data accuracy acquisitions) and integration of hybrid wired/wireless monitoring systems are envisaged. Particular attention is devoted to credible applications designed and deployed to live in harsh environments with intelligent and decision making abilities. A sophisticated automatic, adaptive, sustainable and reliable wireless monitoring system for marine environment has been deployed in Queensland, AUS, November 2007 and an advanced solution is under deployment at the Fiji Islands (2014-2015).

Other applications refer to intelligent embedded systems for rockfall collapse where both traditional and novel sensor are considered. Several top-world still-alive deployments differentiating in the sensor platforms and considered technology are still active and spread out between Italy and Switzerland to monitor catastrophic events as those induced by rockfalls and land slides. Rockfall monitoring: S.Martino Mountain, April 2010 (I); Torrioni di Rialba, July 2010 (I); Val Canaria, Canton Ticino, August 2011 (CH); Gallivaggio, July 2012 (I).  Landslide monitoring: Torrioni di Rialba, July 2011(I), Premana, August 2012 (I), Val Canaria, Canton Ticino, September 2012 (CH). Aspects related to intelligent power management, remote units reconfigurability, remote code upload, data security and effective data storage, aggregation and visualization are object of the research.


Application-level analysis, synthesis and diagnosis of embedded systems

The ongoing research addresses application level properties of the computational flow associated with an embedded system and its relationships with low level design aspects. The developed methodologies and theories for analysis, synthesis and diagnosis, based on the theory of learning and randomized algorithms approaches, allow us to fully characterize the nature of the computation with an acceptable complexity.

Such information can be used to measure the robustness/sensitivity of the application (analysis phase) in the large, provide design guidelines (synthesis phase) and detect, identify and isolate faults and malfunctioning in embedded systems (diagnosis phase). A theory about probably approximated correct computation, i.e., a theoretical framework based on machine learning characterizing performances of embedded systems working within a perturbed environment has been developed and is being assessed.