Comparison of Preprocessors for Machine Learning in the Predictive Maintenance Domain
The optimization of maintenance schedules by predicting machine-health-information is a very obvious application domain for Machine Learning techniques. As part of an ongoing research project, we investigate the ability of Machine Learning to contribute to error prediction in noisy measurements using various preprocessing methods. We focus our efforts on drift-based error patterns and chose several well-established approaches from quality-assurance literature as potential preprocessors for three different Machine Learning methods. We compare the performance of these methods in predicting machine failures, using data based on realistic measurements taken from a supply line in an automotive factory. We further expose the tested methods to different levels of noise and evaluate the performance in detail for drift patterns of varying severity.