Data Analysis for Predictive Maintenance of a Straightening Machine in the Steel Industry

The availability of industrial machinery is crucial to any business operating in the manufacturingsector. Mechanical failures halt production and unplanned downtime canbe disruptive and costly. Small failures can compound to serious failures which exponentiallyincreases downtime and repair costs. Therefore Identifying a degradationcondition before reaching failure is key to maintaining machine availability. On theother hand, it’s undesirable to spend resources performing maintenance that is notrequired. For these reasons a large field of academic work is dedicated to analyzingthe health of a machine, it’s remaining life and in turn preventing failures. This thesis analyses data from a tube straightening machine used in the steel industrywith the goal of implementing a condition monitoring strategy. The data comes froma real world application provided by a multinational manufacturer of steel products.It was obtained using the existing sensors and data acquisition system. The projectserves as a study of the existing infrastructure (available sensors) and it’s suitability forimplementing a condition monitoring strategy. The work is the first step in a largerstudy and does not attempt to perform any implementation or fault identification. Ina broad sense the aim of the project is to identify relationships and patterns in the datathat could be varying with time as the machine degrades. The data consists of twelve channels taken over a two week duration. It is prepossessedto isolate periods where the machine is operating and separated into cycles. Each ofthese is then further processed to extract time and frequency domain features. Thefeatures within each channel are compared with each other using the R2 coefficientof determination to find combinations that are correlated. A semi automated processis used to select the feature combinations. The same process is performed betweensignals for each feature. A number of linear regression models are created based on the results from the correlatedfeatures as well as some multivariate models. These are then compared usinga goodness of fit metric, Normalized Root Mean Square Error (NRMSE). Potentialclustering of machine states are highlighted based on observations in the feature combinations.The conclusions drawn from this study include identification of correlationsbetween signals, potential non-linear relationships and suggestions for future data collectionand analysis going forward. No one feature was identified as correlated betweenall signals