@article{SchrammPieperVogl, author = {Schramm, Simon and Pieper, Matthias and Vogl, Stefanie}, title = {Orthogonal Procrustes and Machine Learning: Predicting Bill of Materials errors on time}, series = {Computers \& Industrial Engineering}, volume = {185}, journal = {Computers \& Industrial Engineering}, publisher = {Elsevier}, doi = {10.1016/j.cie.2023.109606}, abstract = {In an industrial product development process, the Bill of Materials (BOM) is a hierarchical, multi-level representation of all components, parts and quantities of a product. With increasing complexity of industrial products, also BOMs become more complex and thus prone to errors, for example when the individual parts of a product are changed during the product development process. Frequently, these Bill of Materials errors have to be identified manually or by using simple, rule-based schemes. In this paper, we provide a technical background of BOMs, showing the intricacy of temporal BOMs errors in an industrial product development process. The work of other authors, which focused on association mining and tree reconciliation to detect Bill of Materials errors, is analysed. We found that there is currently no system being able to prescribe where in a Bill of Materials and when in the product development process, errors are probable to occur. Also, Machine Learning (ML) methods have not been applied yet. Based on these findings, we formalize the notions Bill of Materials and Bill of Materials errors. Furthermore, we present a deterministic distance measure for BOMS. We provide an answer to the main question of how to represent a Bill of Materials for Machine Learning tasks by solving the orthogonal Procrustes problem for dynamic, hierarchical datasets. Then, we describe an isolation forest based approach to temporal anomaly detection, which points at potential errors in a Bill of Materials at a specific timestamp. Furthermore, we apply Machine Learning and present a multi-output Multi Layer Perceptron for the prediction of temporal Bill of Materials errors. The model predicts where and at which point of time Bill of Materials errors are probable to occur, which renders it a prescriptive system. Eventually, we optimize the performance of our model using contextualization via -means clustering. Finally, we apply our prescriptive pipeline to a real world dataset and show its superiority to existing methods using a qualitative comparison.}, language = {en} } @article{SchwabPogrebnojFreundetal., author = {Schwab, E. and Pogrebnoj, S. and Freund, M. and Flossmann, F. and Vogl, Stefanie and Frommolt, K.-H.}, title = {Automated bat call classification using deep convolutional neural networks}, series = {Bioacoustics}, journal = {Bioacoustics}, publisher = {Taylor \& Francis}, doi = {10.1080/09524622.2022.2050816}, pages = {1 -- 16}, abstract = {Identification of bats is most practically done by exploiting the characteristic features of their echolocation calls. This usually involves expert knowledge, expensive equipment and time-consuming post processing of previously recorded calls. Automated solutions exist, but are usually not as accurate as human experts. We present an automated solution for the processing of bat calls and identification of bat species with extremely high classification accuracy that can be used during live recording or in an automated post-processing software. Our algorithm is the first application of a Deep Convolutional Neural Network to classify bat species based on sound spectrogram images of their echolocation calls. We tested several deep CNN architectures including a modified Google Inception and a ResNet50 architecture. The nets were trained on a very large call database consisting of images of snippets of call spectrograms. All our software was developed in the Python programming language and an executable of the software is available on request.}, language = {en} } @article{BeckDovernVogl, author = {Beck, Nico and Dovern, Jonas and Vogl, Stefanie}, title = {Mind the naive forecast! a rigorous evaluation of forecasting models for time series with low predictability}, series = {Applied Intelligence}, volume = {55}, journal = {Applied Intelligence}, number = {6}, publisher = {Springer}, doi = {10.1007/s10489-025-06268-w}, pages = {27}, abstract = {In the field of time series forecasting, numerous machine learning studies have assessed the performance of new methods on highly volatile data from macroeconomics and finance. Unlike in other domains, where models are also compared to simpler statistical or naive baselines, they mostly compare the performance solely relative to other complex models. This approach may lead to limited conclusions and reduce the practical significance of the results, as it overlooks the unpredictability of some highly volatile time series in the datasets used. We apply state-of-the-art methods from time-series econometrics and machine learning, including autoregressive integrated moving average (ARIMA), exponential smoothing (ETS), Bayesian vector autoregressive model (BVAR), long-short term memory neural networks (LSTM), historical consistent neural networks (HCNN), deep vector autoregressive neural networks (DeepVAR), temporal fusion transformers (TFT), and extreme gradient boosting (XGBoost). Our results demonstrate that no method consistently outperforms the naive (no-change) forecast for highly volatile time series from two popular datasets containing exchange rates and stock prices, rendering comparative analysis between complex models less meaningful. In contrast, when applied to more predictable macroeconomic price indices, many of the methods significantly outperform naive forecasts. We find that the performance of machine learning models deteriorates more than that of statistical models for high-volatility time series. This study highlights the critical importance of using appropriate benchmark models, including cost-effective, simple approaches, on datasets that permit meaningful conclusions.}, language = {en} } @article{VoglLauxBialasetal., author = {Vogl, Stefanie and Laux, Patrick and Bialas, Joachim and Reifenberger, Christian}, title = {Modelling Precipitation Intensities from X-Band Radar Measurements Using Artificial Neural Networks—A Feasibility Study for the Bavarian Oberland Region}, series = {water}, volume = {14}, journal = {water}, number = {3}, publisher = {MDPI}, doi = {10.3390/w14030276}, pages = {1 -- 17}, abstract = {Radar data may potentially provide valuable information for precipitation quantification, especially in regions with a sparse network of in situ observations or in regions with complex topography. Therefore, our aim is to conduct a feasibility study to quantify precipitation intensities based on radar measurements and additional meteorological variables. Beyond the well-established Z-R relationship for the quantification, this study employs Artificial Neural Networks (ANNs) in different settings and analyses their performance. For this purpose, the radar data of a station in Upper Bavaria (Germany) is used and analysed for its performance in quantifying in situ observations. More specifically, the effects of time resolution, time offsets in the input data, and meteorological factors on the performance of the ANNs are investigated. It is found that ANNs that use actual reflectivity as only input are outperforming the standard Z-R relationship in reproducing ground precipitation. This is reflected by an increase in correlation between modelled and observed data from 0.67 (Z-R) to 0.78 (ANN) for hourly and 0.61 to 0.86, respectively, for 10 min time resolution. However, the focus of this study was to investigate if model accuracy benefits from additional input features. It is shown that an expansion of the input feature space by using time-lagged reflectivity with lags up to two and additional meteorological variables such as temperature, relative humidity, and sunshine duration significantly increases model performance. Thus, overall, it is shown that a systematic predictor screening and the correspondent extension of the input feature space substantially improves the performance of a simple Neural Network model. For instance, air temperature and relative humidity provide valuable additional input information. It is concluded that model performance is dependent on all three ingredients: time resolution, time lagged information, and additional meteorological input features. Taking all of these into account, the model performance can be optimized to a correlation of 0.9 and minimum model bias of 0.002 between observed and modelled precipitation data even with a simple ANN architecture.}, language = {en} } @article{PaulFigueroaCotorogeaMarinoVogl, author = {Paul Figueroa Cotorogea, B.S. and Marino, Giuseppe and Vogl, Stefanie}, title = {Data driven health monitoring of Peltier modules using machine-learning-methods}, series = {SLAS Technology}, volume = {27}, journal = {SLAS Technology}, number = {5}, publisher = {Elsevier}, issn = {2472-6303}, doi = {10.1016/j.slast.2022.07.002}, pages = {319 -- 326}, abstract = {Thermal cyclers are used to perform polymerase chain reaction runs (PCR runs) and Peltier modules are the key components in these instruments. The demand for thermal cyclers has strongly increased during the COVID-19 pandemic due to the fact that they are important tools used in the research, identification, and diagnosis of the virus. Even though Peltier modules are quite durable, their failure poses a serious threat to the integrity of the instrument, which can lead to plant shutdowns and sample loss. Therefore, it is highly desirable to be able to predict the state of health of Peltier modules and thus reduce downtime. In this paper methods from three sub-categories of supervised machine learning, namely classical methods, ensemble methods and convolutional neural networks, were compared with respect to their ability to detect the state of health of Peltier modules integrated in thermal cyclers. Device-specific data from on-deck thermal cyclers (ODTC®) supplied by INHECO Industrial Heating \& Cooling GmbH (Fig 1), Martinsried, Germany were used as a database for training the models. The purpose of this study was to investigate methods for data-driven condition monitoring with the aim of integrating predictive analytics into future product platforms. The results show that information about the state of health can be extracted from operational data - most importantly current readings - and that convolutional neural networks were the best at producing a generalized model for fault classification.}, language = {en} } @article{PolzGlawionGebissoetal., author = {Polz, Julius and Glawion, Luca and Gebisso, Hiob and Altenstrasser, Lukas and Graf, Maximilian and Kunstmann, Harald and Vogl, Stefanie and Chwala, Christian}, title = {Temporal Super-Resolution, Ground Adjustment, and Advection Correction of Radar Rainfall Using 3-D-Convolutional Neural Networks}, series = {IEEE Transactions on Geoscience and Remote Sensing}, volume = {62}, journal = {IEEE Transactions on Geoscience and Remote Sensing}, doi = {10.1109/TGRS.2024.3371577}, pages = {1 -- 10}, abstract = {Weather radars are highly sophisticated tools for quantitative precipitation estimation (QPE) and provide observations with unmatched spatial representativeness. However, their indirect measurement of precipitation high above ground leads to strong systematic errors compared to direct rain gauge measurements. Additionally, the temporal undersampling from 5-min instantaneous radar measurements requires advection correction. We present ResRadNet, a 3-D-convolutional residual neural network approach, to reduce these errors and, at the same time, increase the temporal resolution of the radar rainfall fields by a 5-min short-range prediction of 1-min time-steps. The network is trained to process spatiotemporal sequences of radar rainfall estimates from a composite product derived from 17 C-band weather radars in Germany. In contrast to previous approaches, we present a method that emphasizes the generation of spatiotemporally consistent and advection-corrected country-wide rainfall maps. Our approach significantly increased the Pearson correlation coefficient (PCC) of the radar product (from 0.63 to 0.74) and decreased the root mean squared error (mse) by 22\% when compared to 247 rain gauges at a 5-min resolution. An additional large-scale comparison to eight years of data from 1138 independent manual daily gauges confirmed that the improvement is robust and transferable to new locations. Overall, our study shows the benefits of using 3-D convolutional neural networks (CNNs) for weather radar rainfall estimation to provide 1-min, ground-adjusted, that is, bias-corrected with respect to on-ground sensors, and advection-corrected radar rainfall estimates.}, language = {en} } @article{BrunnerVogl, author = {Brunner, Philipp and Vogl, Stefanie}, title = {Extracting Product Improvement Insights from Social Media Comments Using Machine Learning: a Case Study in the Automotive Industry}, series = {Machine Learning and Knowledge Extraction}, volume = {8}, journal = {Machine Learning and Knowledge Extraction}, number = {2}, publisher = {MDPI}, doi = {10.3390/make8020042}, pages = {14}, abstract = {This paper presents a scalable machine learning pipeline for extracting actionable, product-related insights from user-generated social media comments. Leveraging sentence embeddings from SBERT and unsupervised clustering (k-Means and agglomerative), the approach structures informal and noisy comments from Instagram and YouTube into topic groups intended to support thematic analysis. A case study on feedback regarding BMW vehicles, comprising more than 26,000 comments, illustrates how the pipeline can reveal recurring user concerns, such as design critiques, usability issues, and technology-related expectations, even in short and unstructured social media comments. The proposed pipeline operates without labeled data or manual annotation, enabling scalable application and transferability across product categories and industries. By transforming large-scale, unstructured consumer feedback into interpretable themes, the pipeline provides product teams with an efficient and structured basis for data-driven product development and improvement.}, language = {en} }