A distributed self-adaptive nonparametric change-detection test for sensor/actuator networks

The prompt detection of faults and, more in general, changes in stationarity in networked systems such as sensor/actuator networks is a key issue to guarantee robustness and adaptability in applications working in real-life environments. Traditional change-detection methods aiming at assessing the stationarity of a data generating process would require a centralized availability of all observations, solution clearly unacceptable when large scale networks are considered and data have local interest. Differently, distributed solutions based on decentralized change-detection tests exploiting information at the unit and cluster level would be a solution. This work suggests a novel distributed change-detection test which operates at two-levels: the first, running on the unit, is particularly reactive in detecting small changes in the process generating the data, whereas the second exploits distributed information at the cluster-level to reduce false positives. Results can be immediately integrated in the machine learning community where adaptive solutions are envisaged to address changes in stationarity of the considered application. A large experimental campaign shows the effectiveness of the approach both on synthetic and real data applications.