Challenges of Machine Failure Prediction with Product Data - A Case Study

Predictive maintenance aims to minimize downtime and component costs by predicting machine failures. However, the implementation of such a failure prediction system poses several challenges. In this study, a case study is conducted using time series data from a real machine. The time series data of the dataset was recorded at irregular intervals due to product data and machine characteristics. This poses a challenge when using deep learning methods for failure prediction, as they require a coherent fixed latent space. Since non-equidistant time series cannot provide this property, one solution is to convert the time series into an equidistant time series. The objectives of this publication are to present the options for converting non-equidistant time series. Additionally, it will explore the utilization of product data for failure prediction. Subsequently, the failure prediction will be tested using selected methods in a case study. In total, seven conversion possibilities were identified, and the failures were predicted using the product data. This study shows that the deep learning model developed can reduce downtime costs by 10% based on product data.