PSI - Issue 38

Denis Chojnacki et al. / Procedia Structural Integrity 38 (2022) 362–371 Author name / Structural Integrity Procedia 00 (2021) 000 – 000

368

7

Fig. 8. Example of a segmentation during a RLDA campaign

3.2. RLDA campaign improvements: Post-Processing While driving conditions and paths are set, filtering out incidental events remain a necessary step before exploiting the measures. For 10 years, video has become a new standard variable in our RLDA campaigns, to take fast and best decisions when confronted to load extrema. That allows a manually supervised labelling of the RLDA measured events. This is the necessary first step towards unsupervised or self-supervised machine learning labelling algorithm. Let’s see an example on classific ation curve Fig. 9.

Fig. 9. Level crossings of cumulated vertical front left wheel effort of all measures after normalization

Each event identified above some internal standard level as our PG referential has to be analyzed very in detail in order to match with Life situations and to decide after an expertise assessment if it has to be considered as a normal or a misuse event. This activity is fore sure quite time-consuming but mandatory for our referential and potential adaptations of our PG test schedules or internal design standards. 3.3. RLDA campaign improvements : Issues to solve As mentioned above, these campaigns aim to determine and compare road loads on diverse road types. Being obtained with a controlled dr iver’s profile, we need to understand and control this bias on client drivers’ severities, overall and on each road type. Moreover, the MPF used was an average mix, based on customer surveys. A more thorough knowledge of the variety of real client runs should be achieved and exploited in this client characterization process. In order to improve these 2 points we have developed some methodologies based on simplified