IoT Services, AI, WSN, and Mobile Networks
SAF: an innovative energy-efficient system for answering queries, detecting similarities and tracking clusters via time series forecasting
with Sam Madden
We proposed an innovative energy-efficient and adaptable sensor-based system, called SAF, for approximately answering user queries at the sink and for detecting anomalies based on time series forecasting (see publications).
SAF relies on an innovative class of lightweight and adaptable models built at sensors, and on a suite of novel algorithms for monitoring and adapting the prediction models, and for detecting data anomalies and node similarities at no additional communication cost. This is obtained using a novel denition of similarity based on data models, and an efficient clustering algorithm that is optimal in the number of clusters.
SAF offers several benefits: it consumes very little energy, provides data guarantees, and dynamically adapts to variations in the environmental conditions and user requirements.
A resource-efficient self-adaptable hierarchical model-based system to monitor at real-time in very large areas
Most data analytics systems works off-line and rely on a centralized approach, which scales poorly to a very large number of devices and high data rate, consumes battery and is vulnerable to communication failures.
I designed an on-line analytics system that is highly scalable for relying on a distributed intelligence built in a hierarchical manner. More precisely, I proposed a hierarchical approach based on composite statistical models, which are built recursively from simple time series models by aggregating similar nodes using SAF model-based clustering scheme. Moreover, this solution offers self-tunable trade-offs between data accuracy and resource utilization (e.g., energy consumption, network bandwidth).
It can be applied to monitor loads in a large area at different aggregate levels, or to enhance control in a microgrid and maximize its local renewable sources, or to dynamically reconfigure microgrids when switching to islanding mode.
Scalable data sharing in vehicular networks
Ensuring consistency guarantees on shared data in highly mobile networks is a challenging problem that is relevant in several network tasks requiring node coordination (e.g., low duty cycle, object tracking, routing assistance).
We studied under which mobility conditions it is possible to ensure strong consistency data guarantees when node paths and speed are unknown (e.g., in vehicular networks), and provided a minimum set of mobility constraints that are satisable also in case of low density networks.
We also proposed a novel class of quorum systems that is provably optimal in terms of communication costs and that does not violate the consistency property in case of high mobility.
Simulation results based on the restricted random waypoint on a city section had indicated the suitability of our proposal also to vehicular networks.
Energy-efficient and accurate time estimation in low-power devices
Bell-Labs and Univ. Pisa
Time synchronization plays a crucial role in real-time monitoring, node coordination, data aggregation, and in many networking tasks. I studied the time synchronization problem from a novel perspective, which is complementary to the well-studied clock synchronization problem and consists of reducing the error growth between consecutive synchronization using clock information (for further information see publications).
I proposed a suite of deterministic and probabilistic protocols of theoretical interest for refining of the optimality bound for external clock synchronization, and practical interest for
conserving energy in long-lived sensor applications;
improving the clock accuracy by a few constant factors;
enhancing the robustness of the clock in case of network partitions.
A general-purpose on-line fault-detection and diagnosis system
with Mani Srivastava, UCLA
The information provided by WSN monitoring systems must be of high quality to be effective. This is a challenging task in low-power WSN since sensors are error-prone, limited in their energy supply, CPU, memory and bandwidth, and sensitive to environmental variations. Moreover, physical phenomena evolve over the time and are seldom distributed homogeneously.
We proposed an on-line model-based fault-detection system, called Inspect, that is able to detect sensor malfunctioning at real-time and distinguish among stuck-at faults, spikes, outliers and miscalibrations by exploiting spatial-temporal correlations.