PUB - Data-Driven Condition Monitoring of a Hydraulic Press Using Supervised Learning and Neural Networks

Over the lifetime of an industrial machine, wear and tear of various components of the system is inevitable. Especially in applications that involve the transmission of high power under rough conditions, as typically present in hydraulics, the health state of machine elements will degrade over time. In order to maintain high machine availability and process quality, maintenance actions are required. However, the task of planning maintenance actions can be challenging for machine operators when a balance must be met between premature downtimes due to maintenance work and downtimes due to decreased process quality or machine damage. To plan maintenance periods efficiently and detect looming failures at an early stage, methods of condition monitoring can be applied. The objective of condition monitoring is to obtain an estimate of the current machine health status through the automated evaluation of operation data gathered from a machine.

Content

Authors: Makansi, Faried; Schmitz, Katharina

The automated evaluation of machine conditions is key for efficient maintenance planning. Data-driven methods have proven to enable the automated mapping of complex patterns in sensor data to the health state of a system. However, generalizable approaches for the development of such solutions in the framework of industrial applications are not established yet. In this contribution, a procedure is presented for the development of data-driven condition monitoring solutions for industrial hydraulics using supervised learning and neural networks. The proposed method involves feature extraction as well as feature selection and is applied on simulated data of a hydraulic press. Different steps of the development process are investigated regarding the design options and their efficacy in fault classification tasks. High classification accuracies could be achieved with the presented approach, whereas different faults are shown to require different configurations of the classification models.