Machine Learning Model for Predicting Failures in the Context of Autonomous Vehicles

In general, failure data is obtained in the automotive industry during the warranty period. If these contain the expression of a service life characteristic for each failure, statements can be made about the reliability and availability of the systems. For this purpose, estimation methods are used to adapt empirical lifetime distributions to theoretical lifetime distributions. By means of the distribution characteristics, a prognosis of the reliability and availability to be expected is also possible beyond the observation time. In addition to the data basis to be investigated, which is available completely in test bench and experimental trials or re-censored in the field, various methods can be used for estimation. In the past, several methods have been used, such as the estimators according to Eckel [1], Kaplan-Meier [2] or the estimators according to Pauli [3]. In general, a field failure is subject to several stresses, which can be described by expressions of several lifetime characteristics. For this application case, which occurs in the automotive industry, the known estimation methods, [1]–[3], cannot be used. In this paper, the necessity of multidimensional estimation methods will be introduced first. Then, a new estimation method for multidimensional metrics is presented. In a further step, mathematical proof is given that the new method can provide realistic results. Finally, two example data sets from bench testing and field are presented. Furthermore, some recommendations for the use of the new method are concretized in this context.