Analysis with Computational Intelligence techniques of acceleration/temperature data for embedded devices.

The use of time series data is pervasive across a range of industrial sectors. They have already become a fundamental tool in many fields, such as economics, for tasks like data cleaning and fraud detection, but they are now also crucial for detecting machine malfunctions, which is the main focus of this work. The understanding of the characteristics of time series data and the types of anomalies that can be identified is a crucial step in the generation of synthetic replications, essential for the evaluation of anomaly detection methodologies in the absence of annotated data. During my research, I collaborated with AROL, a leading company in the field of capping machines, on an ongoing project. I focused on the Nicla Sense ME, a sensor manufactured by Bosch, which was employed to collect data from a rotational and oscillational motor. The data was transmitted via a Bluetooth Low Energy (BLE) connection, and it comprised the typical information collected in industrial environments, such as temperature and acceleration. State-of-the-art methods for detecting anomalies encompass a range of techniques, from relatively straightforward approaches such as averaging or clustering to more sophisticated methodologies including neural networks and Fourier transformations. The employment of more straightforward methodologies is particularly pertinent when taking into account the constraints imposed by the limited memory and processing capabilities of compact sensors. Furthermore, the transmission of data via Bluetooth in an industrial setting presents a number of challenges. It is therefore beneficial to perform any computation that can be conducted directly on the Nicla Sense ME sensor. In this thesis, I provide a detailed explanation of each method, outlining its distinctive characteristics and the parameters that must be calibrated for effective deployment in real industrial contexts. I conducted comprehensive testing of all the methods on the data set I had collected, utilising a range of anomaly types, and also explored the potential benefits of reducing the number of features employed in the analysis. The results of my testing demonstrate that all the methods I evaluated are capable of performing