Refine
Document Type
- Article (4)
- conference proceeding (article) (2)
- Preprint (1)
Language
- English (7)
Has Fulltext
- no (7)
Is part of the Bibliography
- no (7)
Keywords
- Photoacoustic spectroscopy (2)
- Simulation (2)
- 2D materials (1)
- Acetone (1)
- Atemluft (1)
- Atmospheric measurements (1)
- Atmospheric modeling (1)
- Breath analysis (1)
- Cameras (1)
- Cloud computing (1)
Institute
Begutachtungsstatus
- peer-reviewed (5)
In process analytics or environmental monitoring, the real-time recording of the composition of complex samples over a long period of time presents a great challenge. Promising solutions are label-free techniques such as surface plasmon resonance (SPR) spectroscopy. They are, however, often limited due to poor reversibility of analyte binding. In this work, we introduce how SPR imaging in combination with a semi-selective functional surface and smart data analysis can identify small and chemically similar molecules. Our sensor uses individual functional spots made from different ratios of graphene oxide and reduced graphene oxide, which generate a unique signal pattern depending on the analyte due to different binding affinities. These patterns allow four purine bases to be distinguished after classification using a convolutional neural network (CNN) at concentrations as low as 50 μM. The validation and test set classification accuracies were constant across multiple measurements on multiple sensors using a standard CNN, which promises to serve as a future method for developing online sensors in complex mixtures.
Artificial neural networks (ANNs) are used in quantitative infrared gas spectroscopy to predict concentrations on multi-component absorption spectra. Training of ANNs requires vast amounts of labelled training data which may be elaborate and time consuming to obtain. Additional data can be gained by the utilization of synthetically generated spectra, but at the cost of systematic deviations to measured data. Here, we present two approaches to train ANNs with a combination of comparatively small, measured data sets and synthetically generated data. For the first approach a neural network is trained hybridly with synthetically generated infrared absorption spectra of mixtures of N2O and CO and measured zero-gas spectra, taken with a mid-infrared dual comb spectrometer. This improves the mean absolute error (MAE) of the network predictions from 0.46 to 0.01 ppmV and 0.24 to 0.01 ppmV for the concentration predictions of N2O and CO respectively for zero-gas measurements which was previously observed for training with purely synthetic data. At the same time a similar performance on spectra from gas mixtures of 0–100 ppmV N2O and 0 to 60 ppmV CO was achieved. For the second approach an ANN pre-trained on synthetic infrared spectra of mixtures of acetone and ethanol is retrained on a small dataset consisting of 26 spectra taken with a mid-infrared photoacoustic spectrometer. In this case the MAE for the concentration predictions of ethanol and acetone are improved by 45 % and 20 % in comparison to purely synthetic training. This shows the capability of using synthetically generated data to train ANNs in combination with small amounts of measured data to further improve neural networks for gas sensing and the transferability between different sensing approaches.
To decrease the number of kilometers driven during the development of autonomous cars or driving assistance systems, performant simulation tools are necessary. Currently, domain distance effects between simulation and reality are limiting the successful application of rendering engines in data-driven perception tasks. In order to mitigate those domain distance effects, simulation tools have to be as close to reality as possible for the given task. For optical sensors like cameras, the luminance of the scene is essential. We provide within this paper a method to measure the luminance of rendered scenes within CARLA, an often used open-source simulation environment. Thereby, it is possible to validate the environment and weather models by taking real-world measurements with photometric sensors or with the help of open-source weather data, published e.g. by the German federal service for weather data (DWD - "Deutscher Wetterdienst"). Employing our proposed luminance measurement, the domain gap resulting from the simulation can be specified, which makes it possible to evaluate the statements about the safety of the automated driving system determined within the simulation. We show that the ratio between global and diffuse radiation modeled by the default atmosphere models within CARLA are under limited conditions similar to real-world measurements taken by the DWD. Nevertheless, we show, that the ratio’s temporal variability in real-world situations is not modeled by CARLA.
A Deep Learning System to Transform Cross-Section Spectra to Varying Environmental Conditions
(2022)
Absorption cross-sections provide a basis for many gas sensing applications. Therefore, any error in molecular cross-sections caused by varying environmental conditions propagates to spectroscopic applications. Original molecular cross-sections in varying environmental conditions can only be simulated for some molecules, whereas for most multi-atom molecules, one must rely on high-precision measurements at certain environmental configurations. In this study, a deep learning system trained with simulated absorption cross-sections for predicting cross-sections at a different pressure configuration is presented. The system’s capability to transfer to measured, multi-atom cross-sections is demonstrated. Thus, it provides an alternative to (pseudo-) line lists whenever the required information for simulation is unavailable. The predictive performance of the system was evaluated on validation data via simulation, and its transfer learning capabilities were demonstrated on actual measurement chlorine nitrate data. From the comparison between the system and line lists, the system shows slightly worse performance than pseudo-line lists but its predictive quality is still deemed acceptable with less than 5% relative integral change with a highly localized error around the peak center. This opens a promising way for further research to use deep learning to simulate the effect of varying environmental conditions on absorption cross-sections.
Trace gas analysis in breath is challenging due to the vast number of different components. We present a highly sensitive quantum cascade laser based photoacoustic setup for breath analysis. Scanning the range between 8260 and 8270 nm with a spectral resolution of 48 pm, we are able to quantify acetone and ethanol within a typical breath matrix containing water and CO2. We photoacoustically acquired spectra within this region of mid-infra-red light and prove that those spectra do not suffer from non-spectral interferences. The purely additive behavior of a breath sample spectrum was verified by comparing it with the independently acquired single component spectra using Pearson and Spearman correlation coefficients. A previously presented simulation approach is improved and an error attribution study is presented. With a 3σ detection limit of 6.5 ppbV in terms of ethanol and 250 pptV regarding acetone, our system is among the best performing presented so far.
Trace gas analysis in breath is challenging due to the vast number of different components. We present a highly sensitive quantum cascade laser based photoacoustic setup for breath analysis. Scanning the range between 8263 and 8270 nm with a spectral resolution of 48 pm, we are able to quantify acetone and ethanol within a typical breath matrix containing water and CO2. We photoacoustically acquired spectra within this region of mid-infra-red light and prove that those spectra do not suffer from non-spectral interferences. The purely additive behavior of a breath sample spectrum was verified by comparing it with the independently acquired single component spectra using Pearson and Spearman correlation coefficients. A previously presented simulation approach is improved and an error attribution study is presented. With a 3σ detection limit of 6.5 ppbv in terms of ethanol and 250 pptv regarding acetone, our system is among the best performing presented so far.