From raw data to information for a continuous supervision of machinery in dynamic industrial environments: A case study

Prognostic Health Management (PHM) is a recent approach for the realization of Predictive Maintenance. In literature, there are many papers dedicated to that topic, as well as to each of its part, i.e. signal processing, feature extraction, diagnostic and prognostic. However, when approaching to complex systems operating in real industrial contexts, there are several problems that make difficult its application. First, sensors positioned on the machinery couldn’t have been thought for maintenance purposes. From PHM-application point of view, this could lead to a timeconsuming data pre-processing activity, due to (1) a huge unnecessary amount of data, collected at high frequencies and (2) their intermittent collection during the machinery tests. Second, machinery often work under different operating conditions, that may depend on the kind of product or material that they process; they are often spread worldwide, so the same operating condition could be actually implemented in slightly d...

Prognostic Health Management (PHM) is a recent approach for the realization of Predictive Maintenance. In literature, there are many papers dedicated to that topic, as well as to each of its part, i.e. signal processing, feature extraction, diagnostic and prognostic. However, when approaching to complex systems operating in real industrial contexts, there are several problems that make difficult its application. First, sensors positioned on the machinery couldn’t have been thought for maintenance purposes. From PHM-application point of view, this could lead to a timeconsuming data pre-processing activity, due to (1) a huge unnecessary amount of data, collected at high frequencies and (2) their intermittent collection during the machinery tests. Second, machinery often work under different operating conditions, that may depend on the kind of product or material that they process; they are often spread worldwide, so the same operating condition could be actually implemented in slightly different ways based on the geographic area the plant is installed in. Operating conditions may be unknown during the data collection and even if they are known for a specific machinery, they could change for another machinery, resulting in the impossibility to adopt a same feature, or set of features, and a supervised algorithm for diagnostic for the same machinery. In this paper, a methodology for data pre-processing, feature extraction and condition recognition is introduced through a discussion on a real case study. In particular, the data pre-processing takes into consideration both the quantity and the intermittent nature of collected data, by conducting sampling activities, detecting unstable conditions and setting them apart from subsequent classification; feature extraction and class recognition are conducted automatically, adaptively, and in real-time, so to always know the condition under which machinery is operating and ultimately to make easier the real-time anomaly detection and prognostics.