Real-time estimation of temperatures in indoor environments is critical for several reasons, including the upkeep of comfort levels, the fulfillment of legal requirements, and energy efficiency. Unfortunately, setting an adequate number of sensors at the desired locations to ensure a uniform monitoring of the temperature in a given premise may be troublesome. Virtual sensing is a set of techniques to replace a subset of physical sensors by virtual ones, allowing the monitoring of unreachable locations, reducing the sensors deployment costs, and providing a fallback solution for sensor failures. In this paper, we deal with temperature monitoring in an open space office, where a set of physical sensors is deployed at uneven locations. Our main goal is to develop a black-box virtual sensing framework, completely independent of the physical characteristics of the considered scenario, that, in principle, can be adapted to any indoor environment. We first perform a systematic analysis of various distance metrics that can be used to determine the best sensors on which to base temperature monitoring. Then, following a genetic programming approach, we design a novel metric that combines and summarizes information brought by the considered distance metrics, outperforming their effectiveness. Thereafter, we propose a general and automatic approach to the problem of determining the best subset of sensors that are worth keeping in a given room. Leveraging the selected sensors, we then conduct a comprehensive assessment of different strategies for the prediction of temperatures observed by physical sensors based on other sensors’ data, also evaluating the reliability of the generated outputs. The results show that, at least in the given scenario, the proposed black-box approach is capable of automatically selecting a subset of sensors and of deriving a virtual sensing model for an accurate and efficient monitoring of the environment.

In several domains, systems generate continuous streams of data during their execution, including meaningful telemetry information, that can be used to perform tasks like preemptive failure detection. Deep learning models have been exploited for these tasks with increasing success, but they hardly provide guarantees over their execution, a problem which is exacerbated by their lack of interpretability. In many critical contexts, formal methods, which ensure the correct behaviour of a system, are thus necessary. However, specifying in advance all the relevant properties and building a complete model of the system against which to check them is often out of reach in real-world scenarios. To overcome these limitations, we design a framework that resorts to monitoring, a lightweight runtime verification technique that does not require an explicit model specification, and pairs it with machine learning. Its goal is to automatically derive relevant properties, related to a bad behaviour of the considered system, encoded by means of formulas of Signal Temporal Logic (STL). Results based on experiments performed on well-known benchmark datasets show that the proposed framework is able to effectively anticipate critical system behaviours in an online setting, providing human-interpretable results.

Real-time measurement of temperatures in indoor environments is important for several reasons, among which we mention the maintenance of comfort levels, the satisfaction of legal requirements, and the energy efficiency. However, placing a sufficient number of sensors at the required locations to guarantee a uniform monitoring of the temperature in a given premise may be difficult, with the result that typically just one or a few sensors are deployed. This is the case, for instance, with thermostats in buildings. Virtual sensing is a technique by which values from physical sensors are replaced by those obtained from virtual ones, which take readings from real sensors and calculate their outputs by means of some process models. In this paper, we consider the case study of temperature monitoring in an open office at Silicon Austria Labs, in Villach (Austria). We perform a comprehensive evaluation of various techniques for the prediction of temperatures recorded by physical sensors on the basis of other sensors, ranging from simple baseline methodologies to more complex classical machine learning and deep learning approaches that allow one to take into account temporal and spatio-temporal relationships in the data. The outcome is that, in this context, it is possible to reach a satisfactory prediction performance by using relatively simple models.