Machine learning overview

ML algorithms are widely employed in process engineering and often hidden behind the terms of AI , soft sensing , data fusion or digital twin . In the process industry, typically, large amounts of data are available from the physical sensors in a production plant. This data can be collected, combined and processed by means of ML algorithms (data fusion, soft(ware) sensors) to obtain more meaningful data for the training of data-driven models that capture the real process conditions , . The virtual model (digital twin) can be further used to optimize the production process and an implementation in a loop with the physical world ensures the adaptability of the model . Commonly, one of the following three goals is pursued when applying ML techniques in the process industry: (i) online prediction, (ii) process monitoring, (iii) process fault detection . Some of the most popular algorithms to achieve these goals are provided in Table 1 . The ML algorithms are classified into supervised or unsupervised depending on training data being labeled or unlabeled, respectively. The working principles behind those algorithms with some example applications are concisely summarized in Geet al.  .
Table 1: Examples for popular unsupervised and supervised ML methods.