Using Multivariate Split Analysis for an Improved Maintenance of Automotive Diagnosis Functions

The amount of automotive software functions is continuously growing. With their interactions and dependencies increasing, the diagnosis' task of differencing between symptoms indicating a fault, the fault cause itself and uncorrelated data gets enormously difficult and complex. For instance, up to 40% of automotive software functions are contributable to diagnostic functions, resulting in approximately three million lines of diagnostic code. The diagnosis' complexity is additionally increased by legal requirements forcing automotive manufacturers maintaining the diagnosis of their cars for 15 years after the end of the car's series production. Clearly, maintaining these complex functions over such an extend time span is a difficult and tedious task. Since data from diagnosis incidents has been transferred back to the OEMs for some years, analysing this data with statistic techniques promises a huge facilitation of the diagnosis' maintenance. In this paper we use multivariate split analysis to filter diagnosis data for symptoms having real impact on faults and their repair measures, thus detecting diagnosis functions which have to be updated as they contain irrelevant or erroneous observations and/or repair measurements. A key factor for performing an unbiased split analysis is to determine an ideally representative control data set for a given test data set showing some property whose influence is to be studied. In this paper, we present a performant algorithm for creating such a representative control data set out of a very large initial data collection. This approach facilitates the analysis and maintenance of diagnosis functions. It has been successfully evaluated on case studies and is part of BMW's continuous improvement process for automotive diagnosis.