Filtern
Dokumenttyp
- Vortrag (2)
- Beitrag zu einem Sammelband (1)
- Dissertation (1)
Sprache
- Englisch (4)
Referierte Publikation
- nein (4)
Schlagworte
- Neural Networks (4) (entfernen)
Organisationseinheit der BAM
Eingeladener Vortrag
- nein (2)
We present a workflow for obtaining fully trained artificial neural networks that can perform automatic particle segmentations of agglomerated, non-spherical nanoparticles from electron microscopy images “from scratch”, without the need for large training data sets of manually annotated images. This is achieved by using unsupervised learning for most of the training dataset generation, making heavy use of generative adversarial networks and especially unpaired image-to-image translation via cycle-consistent adversarial networks. The whole process only requires about 15 minutes of hands-on time by a user and can typically be finished within less than 12 hours when training on a single graphics card (GPU). After training, SEM image analysis can be carried out by the artificial neural network within seconds, and the segmented images can be used for automatically extracting and calculating various other particle size and shape descriptors.
The present thesis provides a contribution to the solution of the inverse heat conduction problem in welding simulation. The solution strategy is governed by the need that the phenomenological simulation model utilised for the direct solution has to provide calculation results within short computational time. This is a fundamental criterion in order to apply optimisation algorithms for the detection of optimal model parameter sets. The direct simulation model focuses on the application of functional-analytical methods for solving the corresponding partial differential equation of heat conduction. In particular, volume heat sources with a bounding of the domain of action are applied. Besides the known normal and exponential distribution, the models are extended by the introduction of parabolically distributed heat sources. Furthermore, the movement on finite specimens under consideration of curved trajectories has been introduced and solved analytically. The calibration of heat source models against experimental reference data involves the simultaneous adaptation of model parameters. Here, the global parameter space is searched in a randomised manner. However, an optimisation pre-processing is needed to get information about the sensitivity of the weld characteristics like weld pool dimension or objective function due to a change of the model parameters. Because of their low computational cost functional-analytical models are well suited to allow extensive sensitivity studies which is demonstrated in this thesis. For real welding experiments the applicability of the simulation framework to reconstruct the temperature field is shown. In addition, computational experiments are performed that allow to evaluate which experimental reference data is needed to represent the temperature field uniquely. Moreover, the influence of the reference data like fusion line in the cross section or temperature measurements are examined concerning the response behaviour of the objective function and the uniqueness of the optimisation problem. The efficient solution of the inverse problem requires two aspects, namely fast solutions of the direct problem but also a reasonable number of degrees of freedom of the optimization problem. Hence, a method was developed that allows the direct derivation of the energy distribution by means of the fusion line in the cross section, which allows reducing the dimension of the optimisation problem significantly. All conclusions regarding the sensitivity studies and optimisation behaviour are also valid for numerical models for which reason the investigations can be treated as generic.
Researchers study the properties of polymeric membranes - an important building block of batteries – by first measuring its impedance spectra and then determining a so called equivalent electronic circuit (EEC), that is an electronic circuit that approximately reproduces these measurements. Determining EEC is not completely automatized since it includes the search for the topology of the electronic circuit. This search is an inverse problem, since the formulas for determining the impedance spectra given the circuit topology and parameters are known while there are no formulas for the invers direction. To solve this invers problem, the original problem is split into two parts. The first part is to determine the topology of the circuit and the second part is to determine the parameters of the circuit. For the first part a Convolutional Neural Network Classifier is trained on a simulated data set, where the simulator is an implementation of the formulas for determining the impedance spectra given the complete description of the electronic circuit. After the topology is determined, the parameters of the electronic circuit are found by minimizing the error between the observed spectra and the spectra corresponding to these parameters. This minimization is a global optimization.