PSI - Issue 38

Frédéric Kihm et al. / Procedia Structural Integrity 38 (2022) 12–29

13

2

Kihm, Miu, Bonato / Structural Integrity Procedia 00 (2021) 000 – 000

1. Introduction Accurate damage estimation for automotive systems is important for an optimized design of subsystems based on real world usage patterns. Reliable damage estimation requires to gather realistic usage data from several user profiles, on different road conditions. The data is typically local mechanical strain, which represents an extensive instrumentation – strain gages - and is not practical for such large-scale deployments. This paper discusses the use of placeholder signals, such as temperature, pressure or acceleration, instead of strain signals for the purpose of damage estimation. It will be shown that accurate damage estimation is sometimes possible even with a small subset of signals and that increasing the number of signals brings diminishing returns. This paper introduces a methodology for identifying relevant signals and constructing predictive models for estimating damage. This methodology involves skills from both the engineering and the data science worlds, showing how powerful such a combined approach can be. Two case studies illustrate the use of the methodology to construct a predictive model from measured data. 2. Fatigue Damage prediction using a combined engineering and data science approach Failure prediction can be based purely on a data driven model, or solely on a physics driven approach, or on a combination of the two. A data driven method is based on statistical and machine learning techniques applied to the data. This approach may involve only limited knowledge of the product under consideration. There is therefore a need to gather all the available data with no apri ori assumption of which contribute to the failure, and which don’t. Data Scientists will explore the data using advanced mathematical and statistical methods and create a predictive model. A Physics driven approach is based on the knowledge of the product under consideration i.e. its geometry and material characteristics, how it functions, what are its failure modes and what parameters drive it. Engineers can model the physics of the component or system, e.g. by using Finite Element Simulations or mathematical relationships based on physics. The prediction of a failure will be based on a selection of parameters that are thought to be important to monitor because of their supposed contribution to the failure. These parameters will be used as inputs to a physics-based model, which properties may be progressively fine-tuned so the results match more closely with the observations. The following sections describe a step-by-step approach where the skills of both the data scientist and the engineers are combined to model and predict fatigue failure. 2.1. Capturing the data - instrumentation In order to construct a robust fatigue prediction model, the analyst needs a wide range of measurements. Measurements have to take place across a variety of loading scenarios. In the automotive sector, proving grounds are typically used to collect data over a range of test tracks, representing the various conditions the product will evolve in during its real life (Mendez and Dodds (2013)). A “golden vehicle” needs to be instrumented with sensors to capture all the possible inputs to the system and strain gages, positioned as close as possible to the hot spots. The local stress or strain is usually the basis of a fatigue failure prediction (Bannantine (1990), Halfpenny (2001) and Matsuishi and Endo (1968)), so the strain gauges can be considered as the outputs of the system. In practice, example inputs could be vehicle speed, engine speed, temperature close to the hotspot, acceleration to assess shock and vibration data close to the hotspot, force/torque, pressure, strain gauges, etc. The data can be collected from the CAN Bus and/or analogue sensors. CAN Bus data is an inexpensive source of readily available vehicle parameters but may not contain the exact parameter of interest or may not provide measurements with the required accuracy. It is therefore often completed with analogue sensors and traditional data acquisition hardware to collect high precision data at the locations of interest.

UNRESTRICTED