Knife diagnostics with clustering techniques and support vector machines

This paper is about analysis of experimental data, verifying the applicability of signal analysis techniques for condition monitoring of a packaging machine. In particular, the activity focuses on the cutting process that divides a continuous flow of packaging paper into single packages. The cutting process is made by a steel knife driven by a hydraulic system. Actually the knives are frequently substituted, causing frequent stops of the machine and consequent lost production costs. The aim of this paper is developing a diagnostic procedure to assess the wear condition of the blades, reducing the stops for maintenance. The packaging machine was sensorized with pressure sensor that monitors the hydraulic system driving the blade. Processing of the pressure data comprises three main steps: the selection of scalar quantities that could be indicative of the health state of the knife. A clustering analysis to setup a threshold between healthy and faulted knives. Finally, a Support Vector Machine (SVM) model to classify the health state of knife during its lifetime.