The automotive standards and expectations are changing in an unprecedented pace, the regulatory requirements further add to this momentum of change. Increasing complexity, cost pressure, introduction of new technologies achieve a quieter, electric powered, lower emission and autonomous drive. These are the demands which are faced today in the development process, that lead to high complexity in verification and validation procedures of powertrains under continuously changing configurations of the unit under test. Due to the above-mentioned challenges the stability assessment and the efficiency of the testbed operation in testing campaigns of new powertrains is crucial. The standard testing procedure for proving robustness of system parameters and/or configurations allows a diversity of different test programs for the unit under test. Since the test programs can contain different operating conditions, the assessment of the system stability throughout testing is always a challenge. Especially not being able to discover whether the deviations in the monitored signals occur due to the system changes (e.g. calibration) or due to the influence of the changed conditions on the testbed and test equipment as such. Therefore, additional test repetitions, so-called reference point measurements, are performed on a daily basis to ensure the correct operation of the test facility. Needless to say, these further reference point measurements represent a considerable cost and time factor in the overall process. Knowing this situation and the associated drawbacks of the current procedure, the main focus of our analysis is to develop a concept for a new approach to continuously evaluate drifts on the monitored signals for transient testbed measurements while saving time for additional test repetitions and improving the quality of the stability evaluation. The first phase of this new methodology contains a procedure for quality assessment of time series data, including anomaly detection strategies for various signals, with the use of data-driven machine learning methods. The next phase describes the time series segmentation of the measured signals using advanced pattern recognition techniques. The divided segments are then used for the identification of the reoccurring, but significant, transient events and steady-state operation points in the test programs, later referred to as virtual reference points (VRP). Following this, unsupervised learning procedures for clustering are applied to recognize comparable patterns. The last phase comprises the stability analysis of the monitored signals, which is based on predefined KPIs (key performance indicators), and the judgment of significant deviations from the expected historical trend. Finally, the outcome of our methodology allows the derivation of recommendations, for subsequent actions for test facility operators or test engineers.
Dipl.-Ing. Chiara Gei, AVL List, AUSTRIA Dipl.-Ing. Milan Živadinović, AVL List, AUSTRIA