multivariate statistical prediction models for manufacturing time series

Statistical Process Control (SPC) represents one of the most widely adopted methodologies for monitoring industrial processes. Indeed, in modern manufacturing environments, where sensor measurements exhibit strong serial correlation and significant interactions among multiple process variables, it is fundamental to capture out of control, anomalies and temporal multivariate dependencies among them, in order to substantially decrease and avoid the occurrence of false alarms or delay the detection of actual process changes. For this reason, statistical models for multivariate time series have become an essential component of modern SPC systems. In this thesis different statistical methodologies for the analysis of Univariate and Multivariate Time Series are investigated, such as the ARIMA, VARMA and VECM Models. The objective is to evaluate the capability of these models to describe, capture, and forecast the temporal dynamics and mutual dependencies of industrial process variables, with particular focus on Temperature, Pressure, and Vibration measurements largely collected from manufacturing systems. In addition, both Univariate and Multivariate analyses are performed to validate the proposed modeling framework on several datasets. In particular, the Statistical ap- proach is compared with the Deep Learning one by developing the analysis on datasets representing, mainly, Temperature, Pressure and Vibration values. In the end, the obtained results are discussed in terms of model adequacy, predictive quality, computational cost, short/long term prediction and potential application to real-time statistical process monitoring.