Chemie und Prozesstechnik
Filtern
Dokumenttyp
- Vortrag (14)
- Zeitschriftenartikel (12)
- Posterpräsentation (4)
Schlagworte
- Machine learning (30) (entfernen)
Organisationseinheit der BAM
- 6 Materialchemie (10)
- 1 Analytische Chemie; Referenzmaterialien (9)
- 1.6 Anorganische Referenzmaterialien (9)
- 8 Zerstörungsfreie Prüfung (9)
- 6.3 Strukturanalytik (7)
- 8.4 Akustische und elektromagnetische Verfahren (4)
- S Qualitätsinfrastruktur (4)
- S.1 Qualität im Prüfwesen (4)
- 1.1 Anorganische Spurenanalytik (3)
- 1.3 Instrumentelle Analytik (3)
Paper des Monats
- ja (1)
The applications of composite overwrapped pressure vessels (COPVs) in extreme conditions, such as storing hydrogen gases at very high pressure, impose new requirements related to the system's integrity and safety. The development of a structural health monitoring (SHM) system that allows for continuous monitoring of the COPVs provides rich information about the structural integrity of the component. Furthermore, the collected data can be used for different purposes such as increasing the periodic inspection intervals, providing a remaining lifetime prognosis, and also ensuring optimal operating conditions. Ultimately this information can be complementary to the development of the envisioned digital twin of the monitored COPVs. Guided waves (GWs) are preferred to be used in continuous SHM given their ability to travel in complex structures for long distances. However, obtained GW signals are complex and require advanced processing techniques. Machine learning (ML) is increasingly utilized as the main part of the processing pipeline to automatically detect anomalies in the system's integrity. Hence, in this study, we are scrutinizing the potential of using ML to provide continuous monitoring of COPVs based on ultrasonic GW data. Data is collected from a network of sensors consisting of fifteen Piezoelectric (PZT) wafers that were surface mounted on the COPV. Two ML algorithms are used in the automated evaluation procedure (i) a long short-term memory (LSTM) autoencoder for anomaly detection (defects/impact), and (ii) a convolutional neural network (CNN) model for feature extraction and classification of the artificial damage sizes and locations. Additional data augmentation steps are introduced such as modification and addition of random noise to original signals to enhance the model's robustness to uncertainties. Overall, it was shown that the ML algorithms used were able to detect and classify the simulated damage with high accuracy.
Laser powder bed fusion is one of the most promising additive manufacturing techniques for printing complex-shaped metal components. However, the formation of subsurface porosity poses a significant risk to the service lifetime of the printed parts. In-situ monitoring offers the possibility to detect porosity already during manufacturing. Thereby, process feedback control or a manual process interruption to cut financial losses is enabled.
Short-wave infrared thermography can monitor the thermal history of manufactured parts which is closely connected to the probability of porosity formation. Artificial intelligence methods are increasingly used for porosity prediction from the obtained large amounts of complex monitoring data. In this study, we aim to identify the potential and the challenges of deep-learning-assisted porosity prediction based on thermographic in-situ monitoring.
Therefore, the porosity prediction task is studied in detail using an exemplary dataset from the manufacturing of two Haynes282 cuboid components. Our trained 1D convolutional neural network model shows high performance (R² score of 0.90) for the prediction of local porosity in discrete sub-volumes with dimensions of (700 x 700 x 40) μm³.
It could be demonstrated that the regressor correctly predicts layer-wise porosity changes but presumably has limited capability to predict differences in local porosity. Furthermore, there is a need to study the significance of the used thermogram feature inputs to streamline the model and to adjust the monitoring hardware. Moreover, we identified multiple sources of data uncertainty resulting from the in-situ monitoring setup, the registration with the ground truth X-ray-computed tomography data and the used pre-processing workflow that might influence the model’s performance detrimentally.
The prediction of porosity is a crucial task for metal based additive manufacturing techniques such as laser powder bed fusion. Short wave infrared thermography as an in-situ monitoring tool enables the measurement of the surface radiosity during the laser exposure. Based on the thermogram data, the thermal history of the component can be reconstructed which is closely related to the resulting mechanical properties and to the formation of porosity in the part. In this study, we present a novel framework for the local prediction of porosity based on extracted features from thermogram data. The framework consists of a data pre-processing workflow and a supervised deep learning classifier architecture. The data pre-processing
workflow generates samples from thermogram feature data by including feature information from multiple subsequent layers.
Thereby, the prediction of the occurrence of complex process phenomena such as keyhole pores is enabled. A custom convolutional neural network model is used for classification. Themodel is trained and tested on a dataset from thermographic in-situ monitoring of the manufacturing of an AISI 316L stainless steel test component. The impact of the pre-processing parameters and the local void distribution on the classification performance is studied in detail. The presented model achieves an accuracy of 0.96 and an f1-Score of 0.86 for predicting keyhole porosity in small sub-volumes with a dimension of (700 × 700 × 50) μm3. Furthermore, we show that pre-processing parameters such as the porosity threshold for sample
labeling and the number of included subsequent layers are influential for the model performance. Moreover, the model prediction is shown to be sensitive to local porosity changes although it is trained on binary labeled data that disregards the actual sample porosity.
Non-destructive testing of rail tracks is carried out by using rail inspection cars equipped with ultrasonic and eddy current measurement. The evaluation of test data is mainly done manually, supported by a software tool which pre-selects relevant indications shown to the evaluators. The resulting indications have to be checked on-site using hand-held testing equipment. Maintenance interventions are then derived on the basis of these on-site findings.
Overall aim of the AIFRI (Artificial Intelligence For Rail Inspection) project - funded by the German Federal Ministry of Digital and Transport (BMDV) as part of the mFUND programme under funding code 19FS2014 – is to increase the degree of automation of the inspection process from the evaluation of the data to the planning of maintenance interventions. The accuracy of defect detection shall be increased by applying AI methods in order to enable an automated classification of detected indications into risk classes. For this purpose, data from both eddy current inspections and ultrasonic inspections will be used in combination.
Within the framework of this data-driven project, relevant defect patterns and artefacts present in the rail are analysed and implemented into a configurable digital twin. With the help of this digital twin virtual defects can be generated and used to train AI algorithms for detection and classification. With the help of reliability assessment trained AI algorithms will be evaluated with regard to the resulting quality in defect detection and characterisation.
A particular aspect of the development of AI methods is the data fusion of different NDT data sources: Thereby, synergies are used that arise from linking eddy current and ultrasonic inspection data in a combined model.
In the course of the project a demonstrator consisting of the developed IT-tool and an asset management system will be implemented and tested in the field using real-world data.
Applying data-driven AI systems makes it possible to extract patterns from given data, generate predictions and helps making decisions. Material research and testing holds a plethora of AI-based applications, for example, for the automatized search and synthesis of new materials, the detection of materials defects, or the prediction of process and materials parameters (inverse problems). However, AI algorithms can often only be as good as the training data from which the corresponding models are learned. Therefore, it is also indispensable to develop measures for the standardization and quality assurance of such data.
For this purpose, we develop and implement methods from transferring data from various sources into a homogeneous data repository with uniform data descriptions. Through the standardization and corresponding machine-readable interfaces, research data can be made usable and reusable for further data analyses. In addition to the technical implementation of integrative platforms, it is crucial that quality-assured research data management is recognized and implemented as an integral part of daily scientific work. Finally, we provide a vision of how the Federal Institute for Materials Research and Testing can benefit from data-driven AI systems. We discuss early applications and take a peek at future research.
In this study, we explored the detection of weak bonds (WBs) due to contamination and faulty curing (FC) using linear ultrasound and machine learning. For this purpose, aluminium single-lap adhesive joints containing three variants of bonding quality were investigated: perfect bond, WB due to release agent (RA) contamination, and WB due to FC. The data, according to the deviation of the bonding protocol, were arranged in two groups, creating two datasets: distinct and complete. Each dataset included all bonding conditions (perfect, RA, and FC), although the distinct dataset contained only marginal cases, which were expected to be well separable, whereas the complete dataset included data with minor deviations from the bonding protocol. Pulse-echo C-scan images were acquired for all prepared samples in the immersion tank, and 45 features were initially extracted from the time traces representing each bonding group. The initial data were analysed via a t-test and pairwise correlation analysis to reveal statistically significant features. Then, we performed dimensionality reduction using tree-based, recursive, sequential, and linear discriminant analysis (LDA) feature selectors to explore feature importance and classification accuracy with different feature subsets. Finally, the important features identified with the different feature selectors were fed to support vector machine (SVM) classifiers, and the classification accuracies were compared amongst the different feature subsets. The classification accuracy using a distinct dataset in some cases demonstrated nearly 99% accuracy, indicating that significant bonding protocol deviations could be easily detected. It was demonstrated that classification accuracy increased with the number of features. However, even in the case of the 2D feature space obtained using linear discriminant analysis, the bonding quality classification accuracy remained higher than 84%. The feature subspace reduction with LDA demonstrated sufficient classification accuracy and an improvement of nearly 40% in training time compared with that for the initial feature set. Thus, the classical ultrasonic pulse-echo C-scan with an LDA feature transformation and SVM classifier could be used to identify the deviations in the bonding protocol in aluminium single-lap adhesive joints.
Glow discharge optical emission spectroscopy (GD-OES) is a technique for the analysis of solids such as metals, semiconductors, and ceramics. A low-pressure glow discharge plasma is applied in this system, which ‘sputters’ and promotes the sample atoms to a higher energy state. When the atoms return to their ground state, they emit light with characteristic wavelengths, which a spectrometer can detect. Thus, GD-OES combines the advantages of ICP-OES with solid sampling techniques, which enables it to determine the bulk elemental composition and depth profiles. However, direct solid sampling methods such as glow-discharge spectroscopy require reference materials for calibration due to the strong matrix effect.
Reference materials are essential when the accuracy and reliability of measurement results need to be guaranteed to generate confidence in the analysis. These materials are frequently used to determine measurement uncertainty, validate methods, suitability testing, and quality assurance. In addition, they guarantee that measurement results can be compared to recognized reference values. Unfortunately, the availability of certified reference materials suited to calibrate all elements in different matrix materials is limited. Therefore various calibration strategies and the preparation of traceable matrix-matched calibration standards will be discussed.
Machine learning is an essential component of the growing field of data science. Through statistical methods, algorithms are trained to make classifications or predictions, uncovering key insights within data mining projects. Therefore, it was tried in our work to combine GD-OES with machine learning strategies to establish a new and robust calibration model, which can be used to identify the elemental composition and concentration of metals from a single spectrum. For this purpose, copper reference materials from different manufacturers, which contain various impurity elements, were investigated using GD-OES. The obtained spectra information are evaluated with different algorithms (e.g., gradient boosting and artificial neural networks), and the results are compared and discussed in detail.
High-resolution absorption isotopic spectrometry as a tool for aging studies of Li-ion batteries
(2022)
Lithium (Li) is the key element in the manufacturing of batteries. Isotopic study of Li may help to identify the causes of battery aging due to isotopic fractionation during charge/discharge cycles.
Isotopic ratio determination is based on monitoring the isotopic components of lithium by their spin-orbit coupling and its isotopic shift of about 15 pm for the 22P←22S electronic transition around 670.788 nm. In this work, we propose improvements to our previous work [1] by using a higher-resolution double echelle modular spectrometer (HR-DEMON II) coupled to a continuum source graphite furnace atomic absorption spectrometer (HR-CS-GF-AAS) for the isotopic analysis of Li.
The data analysis was carried out by using a decision-tree-based ensemble machine learning (ML) algorithm (XGBoost). A set of samples with 6Li isotope amount fractions ranging from 0.0004 to 0.99 mol mol-1 was used for the algorithm's training. Subsequently, the procedure was validated by a set of stock chemicals (Li2CO3, LiNO3, LiCl, and LiOH) and a BAM candidate reference material, a cathode material (NMC111). Finally, the ML model was applied to determine the isotope ratio of geological samples, including anorthosite, granite, soil, rhyolite, nepheline syenite, and basalt and battery samples. These samples were measured as digested without any further purification step.
Improvements in the optical resolution resolve the lithium isotopic components of the atomic spectra. In the studied geological samples, were found δ7Li values between -0.5 and 4.5 ‰ with a precision range of 1 to 2 ‰. In addition, the proposed method was validated with multi-collector inductively coupled plasma mass spectrometry (MC-ICP-MS), and these results are comparable and compatible.
Learned block iterative shrinkage thresholding algorithm for photothermal super resolution imaging
(2022)
Block-sparse regularization is already well known in active thermal imaging and is used for multiple-measurement-based inverse problems. The main bottleneck of this method is the choice of regularization parameters which differs for each experiment. We show the benefits of using a learned block iterative shrinkage thresholding algorithm (LBISTA) that is able to learn the choice of regularization parameters, without the need to manually select them. In addition, LBISTA enables the determination of a suitable weight matrix to solve the underlying inverse problem. Therefore, in this paper we present LBISTA and compare it with state-of-the-art block iterative shrinkage thresholding using synthetically generated and experimental test data from active thermography for defect reconstruction. Our results show that the use of the learned block-sparse optimization approach provides smaller normalized mean square errors for a small fixed number of iterations. Thus, this allows us to improve the convergence speed and only needs a few iterations to generate accurate defect reconstruction in photothermal super-resolution imaging.
Mass spectrometry-based proteomics provides a holistic snapshot of the entire protein set of living cells on a molecular level. Currently, only a few deep learning approaches exist that involve peptide fragmentation spectra, which represent partial sequence information of proteins. Commonly, these approaches lack the ability to characterize less studied or even unknown patterns in spectra because of their use of explicit domain knowledge. Here, to elevate unrestricted learning from spectra, we introduce ‘ad hoc learning of fragmentation’ (AHLF), a deep learning model that is end-to-end trained on 19.2 million spectra from several phosphoproteomic datasets. AHLF is interpretable, and we show that peak-level feature importance values and pairwise interactions between peaks are in line with corresponding peptide fragments. We demonstrate our approach by detecting post-translational modifications, specifically protein phosphorylation based on only the fragmentation spectrum without a database search. AHLF increases the area under the receiver operating characteristic curve (AUC) by an average of 9.4% on recent phosphoproteomic data compared with the current state of the art on this task. Furthermore, use of AHLF in rescoring search results increases the number of phosphopeptide identifications by a margin of up to 15.1% at a constant false discovery rate. To show the broad applicability of AHLF, we use transfer learning to also detect cross-linked peptides, as used in protein structure analysis, with an AUC of up to 94%.