Research database

A typical scenario for nowadays massive acquisition systems can be modelled as a large number of sensing units, each transforming some physical unknown quantity into samples of random processes that are then transmitted to a central unit collecting them. To reduce transmission bitrate, signals are often compressed by a lossy mechanism that, if properly tuned, is capable of preserving useful information. Having such large amount of data makes also possible the design of a suitable mechanism for the identification of piece of signal deviating from what is usually observed, i.e., the design of data-driven tools and methods in the framework of anomaly detections. Such a scenario can be declined in many important application areas. Among these, project DECORI focus on structural health monitoring of civil infrastructures and on satellites diagnostic. For the processing of the signal necessary for anomaly detection, we consider two different setting where i) the signal is also processed on-board of (or close to) the sensors acquiring it, or ii) it is first dispatched and then processed on a remote node. In case i) computational resources are a fundamental constraint, but the simultaneous availability of both the original and compressed signal make it possible to detect anomalous cases by assessing situations were the compression algorithm performs poorly. Conversely, case ii) is not affected by any significant constraint in terms of computational budget or memory footprint, but processing can be performed only on the compressed data so that the effectiveness of any possible method for anomaly detection is strictly connected to the amount of information extractable from the received data, i.e., it depends on the capability of the on-board compression stage to correctly preserve all useful information. DECORI aims to advance in several directions. The first step forward will be theoretical, addressing the problem on how compression affects distinguishability which has only been minimally investigated in the literature. Answering to this question will allow to optimize compression not only with respect to data rate but also with respect to anomaly detection. Such theoretical framework will be exploited in the design of optimized anomaly detectors based on machine learning/artificial intelligence for both scenarios of local and remote processing. Finally, these methods will be applied to two important use cases, namely structural health monitoring for civil infrastructures and on-board diagnostic of satellites. For both scenarios measured data will be used and made available, as well as the code developed in DECORI to ensure, according to the FAIR principle, complete reusability of the results of the project.