A Comparison Between Supervised Learning Techniques for Predictive Maintenance in Twin Screw Air Compressors

In this work, a comparison between supervised learning techniques is provided for predictive maintenance in twin screw air compressors. Significant data are selected, acquired and stored in an Industry 4.0 context. Different operating conditions of the process are considered. Subsequently to data collection, data analysis and preprocessing phases are performed in order to prepare tailored datasets to be entered into supervised learning classifiers for predictive maintenance. Four classes associated to the required time priority for maintenance are defined, concerning the state of degradation of the oil used by the compressor. In order to compare different supervised learning techniques, also correlation matrix and Principal Component Analysis are exploited. Accuracy, specificity and sensitivity are evaluated together with the confusion matrix. Significant results are obtained which prove that predictive maintenance policies can be applied to twin screw compressors instead of the widely adopted periodic and corrective maintenance policies.