Modeling Data Quality Using Artificial Neural Networks

Managing data quality is an important issue in all technical fields of applications. Demands on quality-assured data in combination with a more diversified quality description are rising with increasing complexity and automation of processes, for instance within advanced driver assistance systems (ADAS). Therefore it is important to use a comprehensive quality model and furthermore to manage and describe data quality throughout processes or sub-processes. This paper focuses on the modeling of data quality in processes which are in general not known in detail or which are too complex to describe all influences on data quality. As emerged during research, artificial neural networks (ANN) are capable for modeling data quality parameters within processes with respect to their interconnections. Since multi-layer feed-forward ANN are required for this task, a large number of examples, depending on the number of quality parameters to be taken into account, is necessary for the supervised learning of the ANN, respectively determining all parameters defining the net. Therefore the general usability of ANN was firstly evaluated for a simple geodetic application, the polar survey, where an unlimited number of learning examples could be generated easily. As will be shown, the quality parameters describing accuracy, availability, completeness and consistency of the data can be modeled using ANN. A combined evaluation of availability, completeness or consistency and accuracy was tested as well. Standard deviations of new points can be determined using ANN with sub-mm accuracy in all cases. To benchmark the usability of ANN for a real practical problem, the complex process of mobile radio location and determination of driver trajectories on the digital road network based on these data, was used. The quality of calculated trajectories could be predicted sufficiently from a number of relevant input parameters such as antenna density and road density. The cross-deviation as an important quality parameter for the trajectories could be predicted with an accuracy of better than 40 m.