Quality classification methods for ball nut assemblies in a multi-view setting

Abstract

The increase in quality standards in the automotive industry requires specifications to be propagated across the supply chain, a challenge exacerbated in domains such as acoustics, where the quality can be subjective. In a cooperation with ThyssenKrupp Presta AG, we quantify the vibroacoustical influence of ball nut assemblies (BNA) on steering gear by iteratively maximizing the correlation of their order spectra under orthogonality constraints in the canonical correlation analysis (CCA) framework. We specifically propose the fusion of both information sources and the consideration of the setting as a multi-view problem, with the steering gear order spectra as a noisy view of the BNA order spectra. We further show that the performance metric of the binary BNA vibroacoustic quality classification can be improved with respect to the currently used method by employing new, machine learning based classification methods which make use of the maximally correlated CCA encodings of the BNA. The fusion of an additional, up to date unusable information source, namely unlabeled data from scrapped BNA, allows the solving of the resulting semi-supervised learning problem, further increasing the BNA vibroacoustic quality classification performance.