Analytische Chemie
Filtern
Dokumenttyp
- Vortrag (3)
- Zeitschriftenartikel (1)
- Posterpräsentation (1)
Sprache
- Englisch (5)
Schlagworte
- Artificial intelligence (5) (entfernen)
Organisationseinheit der BAM
Paper des Monats
- ja (1)
Eingeladener Vortrag
- nein (3)
In X-ray fluorescence (XRF), a sample is excited with X-rays, and the resulting characteristic radiation is detected to detect elements quantitatively and qualitatively. Quantification is traditionally done in several steps:
1. Normalization of the data
2. Determination of the existing elements
3. Fit of the measured spectrum
4. Calculation of concentrations with fundamental parameters / MC simulations / standard based
The problem with standard based procedures is the availability of corresponding standards. The problem with the calculations is that the measured intensities for XRF measurements are matrix-dependent. Calculations must, therefore, be performed iteratively (= time consuming) in order to determine the chemical composition.
First experiments with gold samples have shown the feasibility of machine learning based quantification in principle. A large number of compositions were simulated (> 10000) and analyzed with a deep learning network. For first experiments, an ANN (Artificial Neural Network) with 3 hidden layers and 33x33x33 neurons was used. This network learned the mapping of spectra to concentrations using supervised learning by multidimensional regression. The input layer was formed by the normalized spectrum, and the output layer directly yielded the searched values. The applicability for real samples was shown by measurements on certified reference materials.
In this talk I’ll describe the use of artificial neural networks (ANN) for quantifying X-ray fluorescence (XRF) measurements. The main idea of this talk is to give an overview of the process needed to generate a model that can then be applied to a specific problem.
In XRF, a sample is excited with X-rays and the resulting characteristic radiation is detected to determine elements quantitatively and qualitatively. This is traditionally done in several time-consuming steps. I’ll show the possibilities and problems of using a neural network to realise a "one-click" quantification. This includes generating training data using Monte Carlo simulation and augmenting the existing data set with an ANN to generate more data. The search for the optimal hyperparameters, manually and automatically, is also described. For the case presented, we were able to train a network with a mean absolute error of 0.1% by weight for the synthetic data and 0.7% by weight for a set of experimental data obtained with certified reference materials.
We present a workflow for obtaining fully trained artificial neural networks that can perform automatic particle segmentations of agglomerated, non-spherical nanoparticles from scanning electron microscopy images “from scratch”, without the need for large training data sets of manually annotated images. 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. 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. We compare the segmentation masks obtained with our suggested workflow qualitatively and quantitatively to state-of-the-art methods using various metrics. Finally, we used the segmentation masks for automatically extracting particle size distributions from the SEM images of TiO2 particles, which were in excellent agreement with particle size distributions obtained manually but could be obtained in a fraction of the time.
The rails of modern railways face an enormous wear and tear from ever increasing train speeds and loads. This necessitates diligent non-destructive testing for defects of the entire railway system.
Non-destructive testing of rail tracks is carried out by rail inspection trains equipped with ultrasonic and eddy current test devices. However, the evaluation of the gathered data is mainly done manually with a strong focus on ultrasonic data, and defects are checked on-site using hand-held testing equipment. Maintenance measures are derived based on these on-site findings.
The aim of the AIFRI project (Artificial Intelligence For Rail Inspection) is to
- increase the degree of automation of the inspection process, from the evaluation of the data to the planning of maintenance measures,
- increase the accuracy of defect detection,
- automatically classify detected indications into risk classes.
These aims will be achieved by training a neural network for defect detection and classification. Since the current testing data is unbalanced, insufficiently labeled and largely unverified we will supplement fused, simulated eddy current and ultrasonic testing data in form of a configurable digital twin.
The deployment of machine learning (ML) and deep learning (DL) in structural health monitoring (SHM) faces multiple challenges. Foremost among these is the insufficient availability of extensive high-quality data sets essential for robust training. Within SHM, high-quality data is defined by its accuracy, relevance, and fidelity in representing real-world structural scenarios (pristine as well as damaged). Although methods like data augmentation and creating synthetic data can add to datasets, they frequently sacrifice the authenticity and true representation of the data. Sharing real-world data encapsulating true structural and anomalous scenarios offers promise. However, entities are often reluctant to share raw data, given the potential extraction of sensitive information, leading to trust issues among collaborating entities.
Our study introduces a novel methodology leveraging Federated Learning (FL) to navigate these challenges. Within the FL framework, models are trained in a decentralized manner across different entities, preserving data privacy. In our research, we simulated several scenarios and compared them to traditional local training methods. Employing guided wave (GW) datasets, we distributed the data among different parties (clients) using IID (independent, identically distributed or in other words, statistically identical) mini batches of dataset, as well as non-IID configurations. This approach mirrors real-world data distribution among varied entities, such as hydrogen refueling stations.
In our methodology, the initial round involves individualized training for each client using their unique datasets . Subsequently, the model parameters are sent to the FL server, where they are averaged to construct a global model. In the second round, this global model is disseminated back to the clients to aid in predictive tasks. This iterative process continues for several rounds to reach convergence.
Our findings distinctly highlight the advantages of FL over localized training, evidenced by a marked improvement in prediction accuracy . This research underscores the potential of FL in GW-based SHM, offering a remedy to similar challenges tied to data scarcity in other SHM approaches and paving the way for a new era of collaborative, data-centric monitoring systems.