Early detection of damage in aircraft structures using Machine Learning and FEM-based methods (#82)
A. Cugniere1, O. Tusch1, C. Stönner1, I. Wieser1, S. Belkner1, A. Mösenbacher1
1 IABG mbH, TAM3 Methodenentwicklung und Betriebsfestigkeit, Ottobrunn, Bavaria, Germany
During their lifespan, aircraft structures are subject to significant loadings that generate cracks and, on the long-term, fatigue damage.
To guarantee the safety of the aircraft, a complete testing of the structure must be carried out prior to the certification. Whole ranges of flight conditions are repeatedly simulated in testing facilities in order to validate the virtual models and to demonstrate the tolerance of the real structure with regard to damage.
To ascertain the presence of cracks in the structure, a limited number of strain gauges are disposed at key positions. These strain gauges provide information about the current state of the structure after a number of load cycles.
Cracks can sometimes appear earlier than expected and still be undetected until the first inspection.
An early detection of these cracks means a better understanding of the underlying mechanisms of the cracks propagation.
For this reason, we investigated a methodology based on Machine Learning and Finite-Element methods to improve the detection and localization of cracks or damage in a structure.
The main idea is to use Machine Learning methods to analyze the data recorded over time by the strain gauges, in order to detect anomalies in the first place.
For instance, clustering analysis allows to identify, in a dataset, groups of data points with similarities. In the same way, it can be used to identify data points that do not match these groups of similar data points and which correspond to anomalies.
Another possibility is to use self-organizing maps (also known as Kohonen maps) or competitive neural networks.
A principal component analysis also provides a way to filter out dominant factors (like loading variations, temperature variations …) and isolate variations in the data that are due to minor effects, such as cracks propagation.
Once an anomaly has been detected and classified as a crack appearance, the next step consists in properly spotting and quantifying the crack(s).
That’s when FEM is brought into play.
To identify cracks using the FEM model of the structure, we applied optimization technics that rely on the measurements performed on the real structure.
The results of such optimization are maps of the structure with prospective crack areas.
Keywords: Machine Learning, FEM, Optimization, Clustering, crack detection
Digital twin as solution for time latency in connected HIL test benches (#78)
R. Bartolozzi1, E. - M. Stelter1, D. Nickel2, C. Schyr2, C. Sültrop3, R. Möller1
1 Fraunhofer Institute for Structural Durability and System Reliability LBF, Assemblies and Systems / Numeric System Analysis, Darmstadt, Hesse, Germany
Nowadays, complex mechatronic systems are standard solutions in products of various technological fields. These are characterised by complex interactions of components from different domains (mechanical, electrical, information technology, etc.). The actual and often highly variable loading of such components in real operating conditions is strongly dependent on their interaction with the rest of the system. The integration at system level requires different competences and functions, which are typically spread out between different partners (OEMs, suppliers and development partners). System integration is carried out when system prototypes are available. How the single components behave together when integrated at system level can only be investigated at a late stage of the development process and problems that may arise often require major design modifications. Moreover, crucial information about the actual loading on components is obtained just at this late point.
There is therefore the need of testing at system level at an earlier stage of the development process. This can be addressed by hardware-in-the-loop (HIL) testing, in which a hardware component is coupled with a real-time simulation of the remaining system that is not yet available as hardware. Within the project TechReaL, funded by the German Federal Ministry for Economic Affairs and Energy (BMWi), an enhanced internet based HIL testing technology was developed. This is based on connecting over the internet different test benches and real-time simulation nodes settled at different locations in a single system HIL testing environment.
In a use case of an automotive electric powertrain, test benches of battery, electric motors and power electronics were connected to each other and to a real-time simulation of the complete vehicle dynamics. A typical value of the main sample time of vehicle dynamics simulation in HIL environments is 1 ms. Due to the time latency of several milliseconds of the internet communication, time consistency between input and output signals cannot be guaranteed.
In this work, a solution based on a digital twin of the tested component is presented to overcome this issue. This makes use of a simulation model of the component, whose parameters are continuously estimated by an identification method. This solution was developed and demonstrated for the battery test bench (see figure). The battery test is carried out at a location, whereas the master node of the connected HIL test is represented by the vehicle simulation, carried out at another location. In a conventional HIL test, the battery test bench should be in closed loop with the vehicle real-time simulation, receiving the current signal as input and sending back the voltage signal. In connected testing, this is not possible within the same time step due to the time latency in the signal communication between the two locations. With the proposed solution, the vehicle model is in closed loop with a linear battery equivalent circuit simulation model, which is run at the same location. In order to have a realistic nonlinear behaviour of the battery model, its parameters are continuously updated by the observer function, which is run at the battery test bench location. This function, which is based on a recursive least square (RLS) method, receives, as input, the current signal, the actual voltage response of the battery and the estimated voltage of the battery model.
The obtained results showed a significant reduction of the error in the voltage signal which is fed back to the vehicle simulation. Due to the time latency, a direct connection of the battery test bench to the vehicle simulation would result in errors up to 15%. With the proposed solution, errors in the 0.1%-range were obtained.
Keywords: Digital twin, HIL, system identification, observer, time latency
About a material model and properties as basis for a digital twin for the fatigue approach (#67)
R. Wagener1, H. Kaufmann2, T. Melz2
1 Fraunhofer Institute for Structural Durability and System Reliability LBF, Component-Related Material Behavior, Darmstadt, Hesse, Germany
Digital twins, as used in the context of Industry 4.0, are based on the numerical image of the material behavior. Consequently, the quality of these digital twins are directly linked to the material model and properties used. The family of the load based concepts is based on linear-elastic material behavior. They are easy to use and the numerical effort involved is quite low, but different Wöhler-curves are required to consider influences, such as mean stress and notch effects on fatigue life and strength. On the other hand, the strain-based concept requires an elastic-plastic stress-strain description and a strain-life curve, which is normally derived from strain-controlled fatigue tests with constant amplitudes under fully reversal push-pull loading conditions. For the consideration of different mean stresses, damage parameters are used.
A digital twin for the fatigue approach must contain a method to handle the effects of service loading conditions and their impact on the material behavior. In order to consider cyclic material behavior in the course of the design process of cyclically loaded components and safety relevant parts, the importance of local strain-based fatigue design approaches have been growing continually. For the damage impact of load-time histories on components such as chassis parts, standard service loads with load amplitudes settled in the high cycle and very high cycle fatigue regime as well as overloads and misuse with load amplitudes from the low cycle fatigue regime have to be considered, in order to perform a reliable fatigue life estimation.
Therefore, various methods have been proposed with regard to the assessment of fatigue strength. Together, they pursue the goal of evaluating the component behavior under service loading, which means randomized variable amplitudes, using cyclic material behavior under constant amplitude loading. Most commonly, the linear damage accumulation rule according to Palmgren-Miner and its modifications are used to accumulate damage for different service load sequences, although non-linear damage accumulation rules have also been developed.
In the majority of cases, the impact on the cyclic material behavior is unconsidered. The cyclic stress-strain curve is derived by constant amplitude test well knowing that the gliding character could influence the stress-strain behavior. That means high amplitudes could affect the stress-strain behavior of following lower amplitudes. Therefore, a material model consisting of the cyclic stress-strain curves derived by Incremental Step Tests and the Fatigue Life Curve will be discussed in order to improve the quality of a digital twin for the fatigue approach.
Keywords: cyclic material behavior, Fatigue Life Curve, digital twin