Applied Machine Learning: Reconstruction of Spectral Data for the Classification of Oil-Quality Levels

Abstract

Inmodern complex systems and machines - e.g., automobiles or constructionvehicles - different versions of a "Condition Based Service" (CBS)are deployed for maintenance and supervision. According to the current state ofthe art, CBS is focusing on monitoring of static factors and rules. In the areaof agricultural machines, these are for example operating hours, kilometersdriven or the number of engine starts. The decision to substitute hydraulic oilis determined on the basis of the factors listed. A data-driven procedure isproposed instead to leverage the decision-making process. Thus, this paperpresents a method to support continuous oil monitoring with the emphasis onartificial intelligence using real-world spectral oil-data. The reconstructionof the spectral data is essential, as a complete spectral analysis for theultraviolet and visible range is not available. Instead, a possibility ofreconstruction by sparse supporting wavelengths through neural networks isproposed and benchmarked by standard interpolation methods. Furthermore, a classificationvia a feed-forward neural network with the conjunction of Dynamic Time Warping(DTW) algorithm for the production of labeled data was developed. Conclusively,the extent to which changes in hyper-parameters (number of hidden layers,number of neurons, weight initialization) affect the accuracy of theclassification results have been investigated.