Predicting condition states, based on displacement data, generated by acceleration sensors on industrial linear vibrating screens through neural networks

Vibrating screens are crucial in the waste and mineral processing industries. However, they often lack comprehensive digital monitoring, which necessitates subjective condition assessments. This study introduces a system developed in cooperation with IFE Aufbereitungstechnik GmbH that provides an objective machine state evaluation using permanently installed acceleration sensors, developed by eSensial Data Science GmbH. Unlike previous research, data for this project were collected from a linear vibrating screen, which is operating in a waste processing plant, introducing uncertainties and occasionally missing data due to sensor damage to the analysis. The study focuses on applying supervised machine learning algorithms to predict the machine's operating condition. In particular, decision trees, multilayer perceptron (MLP) networks, and long-short-term memory (LSTM) networks that were evaluated using classical performance metrics such the MSE and the R2-Score. The models were also tested with respect to missing input data. The MLP network achieved a prediction accuracy of over 90%. Further, it displayed the ability to determine previously unlabeled intermediate states. Additionally, the main cause of prediction errors was identified and a method of handling missing input data was developed.