Writing scientific papers is essential for advancing any given research field, yet undergraduate and graduate students often struggle with this task, facing challenges in clearly presenting their ideas and results. Valuable scientific contributions described in papers that are not well-structured and not easy to follow may not get published. This paper addresses these challenges by proposing a framework called STRaWBERRY, which provides a checklist-based guide to help evaluate individual components of paper drafts against essential quality criteria. Additionally, we propose the use of Large Language Models (LLMs) to automate the assessment based on these criteria. The LLM evaluation process encourages active learning (in the educational sense) and allows the drafts to be iteratively refined through feedback from the LLM. We evaluate the STRaWBERRY framework by its use in lectures and the corresponding outcome: STRaWBERRY has been successfully used in 10 university courses, leading to multiple student pub-lications. Furthermore, we evaluate the LLM-based approach by assessing its accuracy in evaluating a selection of sample papers, demonstrating its potential to supplement and enhance traditional proofreading cycles.

Screw fastening is one of the most important manufacturing processes in the aerospace industry. For example, millions of fasteners are utilized in the Boeing 777 to connect around 3 million different structural components. However, screw fasteners mounted by articulated arm robots in the assembly of aircraft parts are subject to positional uncertainties in both dynamic and static applications. The resulting production errors not only have an impact on cost and time but also lead to safety risks. Despite this strong theoretical and practical relevance, only few research approaches focus on the recognition of defective fasteners in aircraft manufacturing, and the detection of incorrectly assembled aeronautical threaded fasteners remains open. To address these challenges, we identified important time series channels and developed a hard voting ensemble of Categorical Boosting (CatBoost) and Extreme Gradient Boosting (XGBoost) to predict the outcomes of screw fasteners mounted by articulated arm robots in the assembly of aircraft parts. Our model achieved a balanced accuracy of 81.08% on a time series dataset containing both kinematic and dynamic data obtained from a robotic assembly process of aeronautical screwed fasteners. While both CatBoost and XGBoost reached high performance metrics individually, the ensemble approach outperformed each model by utilizing their complementary strengths. Overall, these findings underscore the need for robust machine learning methods to enhance safety standards and reliability in critical assembly processes through advanced techniques.

In robot-assisted minimally invasive surgery (RMIS), the absence of haptic feedback presents a significant challenge for surgeons in accurately gauging the forces applied during procedures. However, obtaining precise force/torque (F/T) information at the surgical site is challenging. One key obstacle is distinguishing external forces from those induced by the cable-actuated kinematics of the surgical tool. We present a novel method to eliminate this interference by employing differential F/T measurement. We utilize two miniature 6-axis F/T sensors, positioned proximally and distally, to counterbalance the undesired forces and torques generated by the cable-driven system. To demonstrate the efficacy of this approach, we developed an experimental cable-actuated forceps with two degrees of freedom. We conducted a series of dynamic tests, attaching various weight configurations to the gripper to simulate external forces ranging from 0.5 N to 1.5 N. Subsequently, we evaluated three measurement methods: raw distal sensor readings, differential compensation, and a multilayer perceptron (MLP) that processes a sliding window of inputs from both sensors and actuators. Differential compensation improves performance by 70% over the distal sensor alone, achieving a root-mean-square error (RMSE) of 0.15 N and 3 mNm across the entire dataset. The MLP yields a further improvement of 90% lower RMSE relative to the distal sensor, achieving 0.05 N and 0.5 mNm on a test subset of the data not used for training.

The increasing share of solar photovoltaic (PV) installations brings new challenges for grid stability, since the electricity generated is variable and depends on solar radiation and other meteorological factors. As previous research on nowcasting PV electricity generation has not simultaneously integrated sky images, temperature data, transfer learning, and data augmentation, we combined these approaches and evaluated various methodologies. Our machine learning models were trained on sky images and PV data derived from the SKIPP’D dataset at Stanford University and the SIRTA atmospheric observatory near Paris, enriched with temperature data from Weather Underground. The results indicate that the transfer learning model trained on augmented sky images from Stanford showed competing overall results and outperformed the current benchmark on sunny days with a root mean squared error of 0.570 on the SKIPP’D dataset. Our approach underscores the complex interactions between various predictive factors and highlights the need for sophisticated methods to enhance the nowcasting of PV power generation. Future research should explore alternative configurations and advanced architectures, such as vision transformer models, to further enhance the prediction accuracy and expand the contribution of machine learning to sustainable and environmentally friendly strategies.

The growing complexity of data science (DS) education calls for interactive and innovative teaching methods to improve student engagement and conceptual understanding of mathematical and algorithmic concepts. Traditional lecture-based instruction often struggles to effectively communicate these topics, highlighting the need for dynamic, exploratory learning tools. This paper investigates whether the use of interactive tools in lectures can improve the comprehension of students learning the functioning of DS algorithms. To address this question, two interactive applications were developed using Shiny for Python. While Shiny for R is well-established in academia, Shiny for Python, a newly stabilized web framework, has received little attention in educational research yet. This study systematically evaluates the potential of Shiny for Python in teaching operations research (OR) and algorithm visualization (AV) through two interactive applications: OptiSense, which focuses on linear programming (LP) and sensitivity analysis, and PathSolver, a Dijkstra AV tool that enables step-by-step exploration of shortest-path problems. Both applications were developed using Design Science Research (DSR) and assessed through undergraduate coursework, comparing their effectiveness against traditional teaching methods. The results indicate that interactive tools enhance student engagement, comprehension, and problem-solving abilities by allowing learners to experiment, visualize, and dynamically analyze mathematical models. This research presents one of the first systematic investigations of Shiny for Python in higher education, demonstrating its potential to reduce cognitive barriers and foster an active learning experience in DS education.

The rapid evolution of football (soccer) analytics has been driven by advances in structured data analysis and recently also by Large Language Models (LLMs). However, existing methods often fail to adapt dynamically to evolving queries and lack contextual richness. This paper presents a novel retrieval augmented generation (RAG) approach tailored to descriptive football analytics, which leverages spatio-temporal and opponent-related data to transform structured event and player data into actionable insights. Our approach was evaluated using a subset from the 2023/24 season of the first German division (1. Bundesliga) over multiple game weeks, achieving an average accuracy of 63.3% in generating responses, setting a first benchmark. In particular, our approach demonstrated strong performance in answering temporal and spatial queries with an accuracy of 70%, while challenges in player-specific queries highlight opportunities for further refinement. These results underscore the potential of RAG to improve decision making for analysts, sports journalists, and potentially coaches by providing dynamic and query-specific insights, paving the way for advanced applications in descriptive sports analytics and interdisciplinary approaches in digital transformation.

Phishing attacks continue to be a major cyber security problem, leading to an increasing number of studies looking at defense strategies. Therefore, we propose an LLM-based phishing detection approach that enhances work by Koide et al. by extending the prompts with URLs, adapting the Chain-of-Thought (CoT) and incorporating additional parameters. Our approach calculates a phishing score, which is used for the classification of websites as either phishing or non-phishing. Subsequently, the results of our research should enable the development of more effective LLM-based phishing detection systems and aim to improve cyber security defenses against this threat.

Text-to-image models, which aim to convert text input into images, have gained popularity partly due to their flex-ibility and user-friendliness. However, there are still weaknesses in the generation of images intended to display emotions, visual text, multiple objects, relative positioning, and attribute binding. This study analyzes the weaknesses of three open-source models: Stable Diffusion v2-1, Openjourney, and Dreamlike Photoreal 2.0. The models are compared based on scores for quality, alignment, and aesthetics. The evaluation is based on (a) the metrics ClipS core, Frechet Inception Distance (FID), and Large-scale Artificial Intelligence Open Network (LAION) and (b) human perception obtained in user surveys. The evaluation revealed that all models show predominantly unsatisfactory performance, and the identified weaknesses were confirmed.

The tremendous advances in Artificial Intelligence (AI) open new opportunities for education, with Intelligent Tutoring Systems (ITS) powered by Generative Artificial Intelligence (GenAI) proving to be a promising prospect. Because of this, our work explores state-of-the-art (SOTA) ITS approaches with the integration of Large Language Models (LLMs) to improve programming education. We investigate whether and how a GenAI-based ITS can effectively support students in learning R programming skills. We measured the performance of three current pairings of LLMs and user interfaces: GPT-3.5 via ChatGPT, PaLM 2 via Google Bard, and GPT-4 via Bing. Therefore, we evaluated the LLMs on four types of problem settings when learning/teaching programming. Our experimental results show that the use of generative AI, specifically LLMs for R programming, is promising, where GPT-3.5 yielded the most satisfactory results. Furthermore, the advantages and limitations of our approach are addressed and revealed. Finally, open research directions towards explainable AI (XAI) and integrated self-assessment are pointed out.

Both robot-assisted surgery (RAS) and future manufacturing systems use telemanipulation setups to enable remote control by surgeons in the operating room and assembly technicians. Precision, intuitive handling, as well as short task execution times have the highest priority. In this paper, we present a novel force-sensing stick and velocity-based online motion generator for a robotic telemanipulator. This custom rigid stick with 6 degrees of freedom (DoF) force/torque (F/T) sensing capabilities is considered for telemanipulation. In a first study, 24 subjects perform three tasks which mimic relevant manipulation maneuvers for industrial assembly and RAS: (1) picking and placing objects, (2) inserting a peg into a hole, and (3) moving the end-effector precisely along a specific pathway. In a second study, three subjects repeat the tasks over a longer period of time in order to assess the learning characteristics and long-term effects on task performance and execution times. For comparison, the same tests are carried out using an off-the-shelf 3 DoF motion-based device. Our results show, that both devices lead to similar performance rates and task execution times over all trials. For the force-sensing stick, subjects report an intuitive and natural response to their force input with no perceptible latency. Conclusions of the long-term study are particularly interesting: within only a few days, task execution times for both input devices can be significantly reduced by 53-69%. The present study builds on previous works of the authors presented at the World Haptics Conference 2023 in Delft [1].

The automatic execution of contact-rich assembly tasks such as peg-in-hole insertion still remains a challenge in industrial manufacturing automation. Deep reinforcement learning (RL) enables agents to learn complex robotic skills, but requires extensive data collection and relatively long execution times when trained online on the physical hardware. In this paper, a robotic setup and RL implementation are presented, which can learn how to successfully perform the peg-in-hole insertion task. The state vector for our force-based learning approach only consists of force/torque (F/T) signals from the robot tooltip without using any position information. We introduce a deep Q-learning (DQN) framework adapted to the task at hand and gather a training data set with a total of 984 peg insertion attempts and 7,884 experiences on the physical setup. Based on this, we apply an offline learning process to improve efficiency by training a large number of policies with different parameter configurations in short time. Finally, the best model configurations are deployed and evaluated on the physical setup reaching a 100% success rate for the insertion task with 0.2 mm clearance.

Food and beverage images are omnipresent on the internet, reflecting a growing trend of people sharing their culi-nary experiences on various social media platforms. This visual abundance of food content presents significant opportunities and challenges for businesses and researchers alike, as they seek to harness the potential of these images' aesthetics for marketing, recommendation systems, and gaining valuable insights into consumer preferences and trends. In this paper we extend the topic's current research state (a) by composing a data set extracted from a real-world recipe data base, labeled by a large number of raters (104 laypersons) to reduce biases and enhance relevance for end-users who are typically no experts; (b) by contrasting the effects of transfer learning using pre-trained ImageNet-models compared to training from scratch and thereby investigating the usefulness of the pre- trained models' general features for aesthetics assessment; (c) by showing how the models can be used as an “Oracle” to isolate the effects of high-level image features on image aesthetics. Our experimental results show that while the test set results on our own data set are similar (approx. 80%) for pre-trained models and models trained from scratch, the pre-trained ImageNet-models have better generalization capabilities with respect to a different data set.

This study evaluates the effectiveness of transfer learning models in industrial surface defect detection using few-shot learning. Surface defect detection is a critical task in various industrial applications, where accurately detecting and classifying defects can improve product quality and increase manufacturing efficiency. However, data scarcity is a considerable challenge: obtaining and labelling defect samples is a costly, time-consuming process and difficult due to their infrequent occurrence. Few-Shot learning aims to effectively train models using only a limited number of labelled samples, thus mitigating the impact of data scarcity. This study compares the performance of transfer learning models pre-trained on three different data sets for few-shot learning in the context of surface defect detection. On the one hand, transfer learning models pre-trained on the ImageNet data set yield the best overall results in terms of accuracy. On the other hand, our results indicate that the DAGM data set, an industrial optical inspection data set which is close to the target domain, is particularly effective for training models to clearly detect surface defects in a few-shot learning scenario.

As the amount of available data in healthcare has increased significantly and only 20% of electronic health record data are in a structured format, data lakes have become a common solution for managing heterogeneous data in the healthcare domain. Nowadays, these are utilized far below their capabilities in medical research. Since previous reviews only partly address data lakes in the healthcare domain, a systematic literature review on this topic is missing. Therefore, this paper provides an overview of applications in the healthcare domain that benefit from data lakes. We review the literature and structure it according to data sources and players, and we identify applications and future research needs of data lakes in the healthcare domain. Overall, it turned out that all players could benefit from the capabilities of data lakes. We found that data lakes are currently not broadly implemented in the field, and the viewpoint of hospital operators and healthcare insurers seems to be an underresearched topic compared to the other players.

In a dynamic world, data streams are continuously generated, which poses immense challenges for machine learning (ML) algorithms to adapt to changing statistical properties that are subject to a non-stationary context. The underlying scenario is defined as concept drift (CD), where changes in the relationship between response and prediction variables (real CD) or a change in input data (virtual CD) are accompanied by a significant degradation in the predictive performance of the models, causing ML models to reach unacceptable levels of system accuracy. In this paper, the state of the art for CD algorithms is analyzed and compared. For this purpose, a systematic literature review was performed. Then, the 10 most popular CD algorithms were extracted from the literature using a newly-developed metric. Subsequently, the algorithms were analyzed and compared with respect to their functionality and limitations. Based on these, the optimization potentials were systematically derived. This work presents a summarized overview of CD algorithms and provides the basis for algorithm optimization in this domain.

In the era of digital transformation, supply chain management faces major challenges induced by the lack of transparency and the evolving industry. In this context, blockchain technology has emerged as a possible answer to the future problems of the supply chain. In this paper, we present a systematic literature review on blockchain in the context of supply chains. The goal of our work is to present the factors and capabilities of blockchain technology that contribute to improving supply chain resilience. We also show which supply chain management factors limit the use of blockchain technology. Based on this, we identify various areas and applications of blockchain technology to support supply chains and highlight current work in the field. From the reviewed literature, we deduce a number of open challenges regarding the application of blockchains in the context of supply chains, e.g. the need (a) to improve the implementation process, (b) to make blockchain more cost-effective, (c) to educate potential users regarding blockchain security aspects, and (d) to further digitize supply chains as part of the digital transformation process.

Small data and big data are distinct approaches to data analysis and utilization in various applications. While big data has been the focus of many research and business efforts for more than ten years, small data is increasingly being recognized as having potential value in certain settings. We systematically review literature and conclude that small data can be indeed valuable in certain scenarios. This paper incorporates the data value perspective of small data within various application areas. For this, we apply the data-information-knowledge-wisdom (DIKW) hierarchy to categorize papers and findings, and discuss the papers from the view point of “data value”. Our review identifies various contexts where small data can be used to create value, such as data pre-processing, classification tasks, anomaly detection, forecasting and decision support. We also highlight industries that may be particularly promising areas for practitioners and researchers focused on small data. In addition, we provide an overview of methods and tools for small data analysis, including statistical techniques, visualization, and machine learning algorithms. Finally, based on our results, we suggest, that further research should focus on small data analysis.

Abstract In clinical practice, algorithmic predictions may seriously jeopardise patients’ health and thus are required to be validated by medical experts before a final clinical decision is met. Towards that aim, there is need to incorporate explainable artificial intelligence techniques into medical research. In the specific field of epileptic seizure detection there are several machine learning algorithms but less methods on explaining them in an interpretable way. Therefore, we introduce XAI4EEG: an application-aware approach for an explainable and hybrid deep learning-based detection of seizures in multivariate EEG time series. In XAI4EEG, we combine deep learning models and domain knowledge on seizure detection, namely (a) frequency bands, (b) location of EEG leads and (c) temporal characteristics. XAI4EEG encompasses EEG data preparation, two deep learning models and our proposed explanation module visualizing feature contributions that are obtained by two SHAP explainers, each explaining the predictions of one of the two models. The resulting visual explanations provide an intuitive identification of decision-relevant regions in the spectral, spatial and temporal EEG dimensions. To evaluate XAI4EEG, we conducted a user study, where users were asked to assess the outputs of XAI4EEG, while working under time constraints, in order to emulate the fact that clinical diagnosis is done - more often than not - under time pressure. We found that the visualizations of our explanation module (1) lead to a substantially lower time for validating the predictions and (2) leverage an increase in interpretability, trust and confidence compared to selected SHAP feature contribution plots.

Abstract Labeling of datasets is an essential task for supervised and semi-supervised machine learning. Model-based active learning and user-based interactive labeling are two complementary strategies for this task. We propose VisGIL which, using visual cues, guides the user in the selection of instances to label based on utility measures deduced from an active learning model. We have implemented the approach and conducted a qualitative and quantitative user study and a think-aloud test. The studies reveal that guidance by visual cues improves the trained model’s accuracy, reduces the time needed to label the dataset, and increases users’ confidence while selecting instances. Furthermore, we gained insights regarding how guidance impacts user behavior and how the individual visual cues contribute to user guidance. A video of the approach is available: https://ml-and-vis.org/visgil/ .

In many academic fields, literature review has become an established research method of technical writing. In this process, it serves as a method for identifying relevant findings in a research area by synthesizing existing data, identifying knowledge gaps, and critically evaluating results. We systematically reviewed the literature on writing literature reviews and found that a number of papers on that topic has been published, but they do not include suggestions and guidelines for all typical components of a literature review. Therefore, this paper deals with the research question, what are the typical components of a literature review and what should they contain to achieve a high-quality literature review. This paper first explains the goals of literature reviews and then introduces the most common types of literature reviews. Afterwards, the main components are described, and methodological approaches of different authors are brought in. In addition, the goal-oriented process of individual components is presented. Thereby, the paper does not focus on a specific research area but takes an interdisciplinary approach to the given topic.

Due to its achievements in recent years, machine learning (ML) is now used in a wide variety of domains. Educating ML has hence become an important factor in academia and industry. We argue that students learning about machine learning will need, in addition to theoretical knowledge, approaches to interactively explore Machine Learning models and their parameters. This paper introduces EduML –an interactive approach for lecturers to teach and for students or professionals to study and explore the fundamentals of machine learning. EduML allows users to experiment with data preparation, dimensionality reduction and a wide range of classifiers on different data sets. These data sets can be analysed in order to understand the complexity of the classification problem. The classifiers can be autonomously fitted to the training data or the effect of manually altering model hyperparameters can be explored. Additionally, to get started with programming own ML pipelines, Python and R source code of configured ML pipelines can be extracted. EduML has been used in a lecture as an interactive demo or by students in lab sessions. Both scenarios were evaluated with a user survey.

For the manufacturing of miniature force/torque sensors, extreme accuracy is required due to the tiny size of the strain gauges inside the sensors (2×2.5 mm). The current method of manually assembling them by hand is difficult, time-intensive, and error-prone. To improve this, a system to pick up the tiny objects from a plate and place them on elementary cells is being devised using a 6-axis robot arm with custom end-effector and a camera with magnification lens. This paper focuses on the perception module by evaluating methods for detecting tiny and transparent objects and obtaining spatial information from 2D images. Additionally, it considers aspects of the camera-to-robot calibration process, which are necessary to transfer the accuracy of image recognition into the real world. An approach using image segmentation and blob detection is taken, precluding the need for machine learning models. This is possible due to the superb image quality achieved by the sufficiently advanced camera and lighting setup. As a conclusion, we propose a perception module, which is capable of pinpointing strain gauge positions within ±0.1 mm and can also recognize different types of components based on physical dimensions. Our end-to-end approach for automatic pick-and-place operations integrates the perception module, camera-to-robot calibration, and a last-minute correction routine, which ultimately leads to an overall positioning accuracy of ± 0.3 mm.

In minimally invasive surgery (MIS), the reliable detection of hard inclusions in soft tissue is crucial for the success of the intervention. In robot-assisted surgery (RAS) however, limited technologies are available for intracorporeal tissue stiffness assessment due to the lack of force and tactile feedback from the robot tool tip. This paper investigates both, human haptic perception and robotic F/T sensing in similar experimental setups to draw conclusions about the usage of a haptic sensor for teleoperation in RAS. We use a novel 6-axis F/T sensor compact enough to be moved through trocars during RAS interventions and experimentally analyze its performance in a simulated robotic palpation task. Furthermore, we carry out a comprehensive user study$(n=30)$and collect fingertip interaction data to investigate human haptic perception. Results show, that both approaches detect larger bead diameters of 19 mm and 15 mm with high precision and show similar accuracy rates. With regards to interaction forces, subjects on average apply more than 10 times the amount of normal force$(F_{z}=28.8\pm 4.9\mathrm{N})$, which leads to higher accuracy particularly for smaller embedded nodules. The robotic sensing technique, on the contrary, offers distinct advantages by providing more gentle treatments and reducing the risk of tissue damage.

One big challenge regarding the development of highly automated driving (HAD) functions is validation and, in particular, providing proof of the desired functionality in any given scenario. Especially, corner cases, representing atypical, rare scenarios such as unexpected object movements are of high interest and thus must be detected to treat them with special attention. First, this paper presents a taxonomy for corner cases (CC) with focus on HAD. Specifically, so-called systemic corner cases (SCC) are introduced. Next, a feasibility study is presented on how these SCCs can be detected using different Machine Learning (ML) approaches for anomaly detection. We propose to use a hybrid ensemble of a One-Class Support Vector Machine (OCSVM) and a Clustering-Based Local Outlier Factor (CBLOF) incorporating domain knowledge to account for the nature of corner cases in timely correlated scenarios. The underlying data are unlabeled multivariate time series of HAD-system internal variables. Our experiments on both, synthetically generated and representative real-world CC, show that the hybrid ensemble can detect a variety of real corner cases, which allows for promising validation support of HAD functions.

In machine learning, the presumably best model is selected from a variety of model candidates generated by testing different model types, hyperparameters, or feature subsets. The advent of deep learning has made model selection even more challenging due to the huge parameter search space. Relying on a single metric to select the best model does not consider class imbalances or the different costs of misclassifications. We argue that incorporating human knowledge to interactively analyse the per-class errors and class confusions over all model candidates enables a more efficient training process and yields better models for given applications. This paper proposes the model-agnostic approach ConfusionVis which allows to comparatively evaluate and select multi-class classifiers based on their confusion matrices. This contributes to making the models’ results understandable, while treating the models as black boxes. Therefore, we propose a novel method to measure and visualise distances between confusion matrices and an interactive query interface to incorporate all composition levels of class errors. The approach is evaluated in a user study and the applicability is shown by a case study where marine biologists investigate the conservation efforts of baleen whales by classifying whale species in acoustic recordings. ConfusionVis is available online: https://www.ml-and-vis.org/confusionvis.

The human sense of touch allows recognizing a wide set of properties of a grasped object such as weight, shape, hardness, temperature or surface texture. Despite the great importance of haptic sensing for humans, mechatronic end-effectors of humanoid robots and industrial manipulators are rarely endowed with tactile feedback. This is due to a lack of robust force/torque sensors which are compact enough to be integrated in the robot's fingertips. This paper leverages a novel 6-axis force/torque sensor and investigates, how local force/torque sensing at the end-effector fingertip best enables the robot to classify different surface textures. Fingertip measurements of reaction forces and torques are recorded for a total of 21 textures as the robot performs sliding movements similar to those that humans make when exploring textures. After data collection and signal processing, the extracted features are used for texture recognition, utilizing k-nearest neighbor (kNN), decision tree, random forest as well as multi-layer perceptron (MLP) classifiers. Our experimental results show that the concatenated power spectral densities extracted from the force and torque time series are the most discriminative input features enabling the random forest to achieve an average recognition accuracy of 98.8±0.4%.

In the field of predictive maintenance (PdM), machine learning (ML) has gained importance over the last years. Accompanying this development, an increasing number of papers use non-interpretable ML to address PdM problems. While ML has achieved unprecedented performance in recent years, the lack of model explainability or interpretability may manifest itself in a lack of trust. The interpretability of ML models is researched under the terms explainable AI (XAI) and interpretable ML. In this paper, we review publications addressing PdM problems which are motivated by model interpretability. This comprises intrinsically interpretable models and post-hoc explanations. We identify challenges of interpretable ML for PdM, including (1) evaluation of interpretability, (2) the observation that explanation methods explaining black box models may show black box behavior themselves, (3) non-consistent use of terminology, (4) a lack of research for time series data, (5) coverage of explanations, and finally (6) the inclusion of domain knowledge,

The estimation of a system's or a component's remaining useful life (RUL) is considered the most complex task in predictive maintenance, at the same time the most beneficial one. In this brief review paper, we survey the state-of-the-art in machine learning-based RUL prognosis based on research on NASA's C-MAPSS data set. We identify the frequently used models, comparatively evaluate model performance and survey the used feature extraction methods. As a main contribution, we formulate challenges in the field, independently of the C-MAPSS data set. Among the challenges are interpretability, model uncertainty and domain adaptation, i.e. transfer learning. The identified challenges may serve to identify potential research directions, in order to further push the field of machine learning applied to RUL prognosis.

Obtaining new information and creating value from present measurements without introducing additional sensors is cost-efficient and mitigates data that is collected and stored by information systems but not used. In electromechanical drive systems, defect states of synchronous motors can be detected based on measurements of the motor current. While there is a tendency to (exclusively) apply Deep Learning models to such problems, we argue that, for appropriate problem settings, alternatives should also be evaluated and at least be used as benchmarks. This paper addresses the question of whether non-Deep Learning methods are competitive to Deep Learning ones for sensorless detection and classification of faults in electromechanical drive systems. For this multi-class classification problem, a systematic evaluation of selected traditional, ensemble, and Deep Learning classifiers is conducted for a data set with one normal state and ten fault states of an electromechanical drive system. In addition to working on the raw input data, the impact of Recursive Feature Elimination is compared to dimensionality reduction with Principal Component Analysis. Accuracy, computational complexity, and engineering effort of the different Machine Learning pipelines are compared. A key finding is that the appropriate combination of feature elimination and Machine Learning model yields high accuracies while allowing to massively reduce the number of features, hence making the detection of fault states less computationally expensive.

One of the major problems of applying supervised machine learning methods in real-world problems is the absence of labeled data. Labeling huge amounts of data is time consuming and cost intensive. Moreover, in many cases, labels can only be assigned by domain experts like medical doctors or engineers, who have little time and do not necessarily have profound machine learning knowledge. In this paper, we propose an efficient interactive cluster-clean-label approach. First, to visualize the potentially huge amount of data, principal component analysis followed by t-SNE projection is applied. On the 2-dimensional representation of the data, HDBSCAN clustering is utilized to identify groups of potentially similar class membership. Subsequently, anomaly detection in form of an autoencoder is applied on each cluster, and instances that are likely to belong to different classes are suggested to the user. The user decides which of these suggested instances to include and restarts the anomaly detection process with the remaining subset of instances. This iterative process is repeated until the user is satisfied with the clusters' purity. Eventually, labels are assigned to the clusters. The approach is evaluated by a user study with 25 participants using the initially unlabeled MNIST data set, where on average users were able to label 91.59% of the data set, with an accuracy of 98.99%. A video showing the approach is available: https://youtu.be/RsLI0dg90qE.

Human-centered machine learning is becoming an emerging field aiming to enable domain experts that do not necessarily have a data science background to make use of machine learning applications. Especially in unsupervised machine learning, e.g. cluster analysis, models cannot be autonomously tuned towards an optimal solution for a given application due to the absence of ground truth like class labels. In cluster analysis, different feature subsets may lead to different clusterings. The identification of the best subset of given features is therefore essential in order to improve the overall clustering performance and to obtain a clustering that is suitable for a given application. To support users in finding an optimal clustering solution, we propose XplainableClusterExplorer, an interactive and explorative approach suitable for feature selection for clustering. In an interactive combination of user and machine learning models, the user is supported by evaluation criteria and visualizations in determining feature subsets and adjusting hyperparameters. For feature subset selection we propose a combination with feature importances from random forests and LIME. Since this requires a supervised setting, the cluster assignments are used as tentative class labels in subsequent step. Our experimental results have shown that this subsequent classification step leveraging calculated feature importances can facilitate feature subset selection and therefore enhance overall clustering performance.

The labeling of datasets is an important task for supervised and semi-supervised machine learning that can be addressed with visual analytics. With model-based active learning and user-based interactive labeling, there are two complementary strategies for this task. We present an approach that combines the strengths of both areas and aims to guide users through model-based recommendations and highlighting in one interface when selecting and labeling instances. For this purpose, an active learning strategy is used to recommend useful instances in addition to the user-based selection of instances. We have implemented the approach and conducted a user survey to research the effects guidance by visual cues has on the users' selection strategies. The proposed approach combines both perspectives in a single interactive visualization to support the user with different degrees of guidance in the selection of instances. Our results of the user survey suggest that user guidance has a positive influence on the users' perceived confidence and difficulty in selecting instances, on their orientation, and on their perceived impression of the models' performance. A video of the approach is available: https://youtu.be/TYPWG85Akn0.

The complexity of vehicles has increased over the last years and will continue to do so. Hence, repairs in repair shops become more and more complex and thereby time-consuming, where time to repair is a competitive factor. During repairs and servicing of vehicles in the independent aftermarket the data read-out using diagnostic testers is transferred back to a common back-end. This “Big Data” contains millions of diagnostic trouble codes (DTCs) and freeze frames, where DTCs point to potential fault causes and freeze frames describe the vehicle's condition when the DTC was stored, e.g. the current engine RPM. This paper proposes an approach to benefit from this “Big Data” in order to (a) acquire knowledge for the development of new vehicles and components and (b) to speed up the time for future fault analyses and repairs. The aim is to detect the vehicle operation modes where a specific fault code was previously not or rarely observed, referred to as novelty detection (anomaly detection). This can point to rare fault causes, which are likely to have a longer time to repair since they are not common. From the field of machine learning, one-class classifiers are applied in two setups: (1) one-class novelty detection, where the data items from all operation modes are combined to one training set and (2) multi-class novelty detection, where an individual classifier is trained for each operation mode. On real data it is shown that novelties can be successfully detected. While it was found that many faults predominantly occur while the vehicle is in a specific operation mode, e.g. when the engine is warm and is in idle, the most interesting novelties were faults when the engine was off or cold. It was found that the multi-class case is superior for the underlying problem using autoSVDD to yield the best results.