State-of-the-art algorithms are reported to be almost perfect at distinguishing the vibrations arising from healthy and damaged machine bearings, according to benchmark datasets at least. However, what about their application to new data? In this paper, we confirm that neural networks for bearing fault detection can be crippled by incorrect hyperparameterisation, and also that the correct hyperparameter settings can change when transitioning to new data. The paper combines multiple methods to explain the behaviour of the hyperparameters of a wide-kernel convolutional neural network and how to set them. Since guidance already exists for generic hyperparameters like minibatch size, we focus on how to set architecture-specific hyperparameters such as the width of the convolutional kernels, a topic which might otherwise be obscure. We reflect different data properties by fusing information from seven different benchmark datasets, and our results show that the kernel size in the first layer in particular is sensitive to changes in the data. Looking deeper, we use manipulated copies of one dataset in an attempt to spot why the kernel size sometimes needs to change. The relevance of sampling rate is studied by using different levels of resampling, and spectral content is studied by increasingly filtering out high frequencies. We find that, contrary to speculation in earlier work, high-frequency noise is not the main reason why a wide kernel is preferable to a narrow kernel. Finally, we conclude by stating clear guidance on how to set the hyperparameters of our neural network architecture to work effectively on new data.

Recent work uncovered potential flaws in attribution or heatmap based saliency methods. A typical flaw is a confirmations bias, where the scores are compared to human expectation. Since measuring the quality of saliency methods is hard due to missing ground truth model reasoning, finding general limitations is also hard. This is further complicated, because masking-based evaluation on complex data can easily introduce a bias, as most methods cannot fully ignore inputs. In this work, we extend our previous analysis on the logical dataset framework ANDOR, where we showed that all analysed saliency methods fail to grasp all needed classification information for all possible scenarios. Specifically, this paper extends our previous work using analysis on more datasets, in order to better understand in which scenarios the saliency methods fail. Further, we apply the Global Coherence Representation as an additional evaluation method in order to enable actual input omission.

The Siamese network architecture has been applied by deep learning practitioners to find similarities between images. In the domain of autonomous driving, this network configuration has recently gained attention for solving the change detection task, which involves identifying changes in a previously known map of a vehicle’s environment. This is vital, as such deviations may compromise the accuracy and reliability of the map, which is essential for the vehicle’s ability to localize itself and navigate effectively. In this paper, we present a set of experiments involving state-of-the-art deep learning architectures based on both convolution (CNN) and attention mechanisms such as AlexNet, GoogLeNet, VGG, ResNet, Vision Transformer, and Shifted Windows Transformer as possible candidates for the feature extractor backbone module in the Siamese architecture to detect changes caused by the disappearance and appearance of construction zones. Also, we evaluate the performance of these architectures using fine-tuning, i. e., initializing the convolutional layers with pre-trained weights. In our experimentation, the best results were obtained using VGG16 (CNN), especially when it was initialized using pre-trained weights from the ImageNet-1K dataset. In particular, VGG16 with an average F1 score of 92% on highway datasets outperformed the baseline residual network composed of ResNet18 convolutions by about 13.5%.

Balancing performance and interpretability in multivariate time series classification is a significant challenge due to data complexity and high dimensionality. This paper introduces PHeatPruner, a method integrating persistent homology and sheaf theory to address these challenges. Persistent homology facilitates the pruning of up to 45 maintaining or enhancing the accuracy of models such as Random Forest, CatBoost, XGBoost, and LightGBM, all without depending on posterior probabilities or supervised optimization algorithms. Concurrently, sheaf theory contributes explanatory vectors that provide deeper insights into the data's structural nuances. The approach was validated using the UEA Archive and a mastitis detection dataset for dairy cows. The results demonstrate that PHeatPruner effectively preserves model accuracy. Furthermore, our results highlight PHeatPruner's key features, i.e. simplifying complex data and offering actionable insights without increasing processing time or complexity. This method bridges the gap between complexity reduction and interpretability, suggesting promising applications in various fields.

For studying intrusion detection data we consider data points referring to individual IP addresses and their connections. We build networks represented by graphs associated with those data points, such that vertices in a graph are constructed to denote the respective IP addresses, with the key property that attacked data points are part of the structure of the network. More precisely, this paper proposes a novel approach using simplicial complexes to model the desired network and the respective intrusions in terms of simplicial attributes, thus generalizing previous graph-based approaches. Applying adapted network centrality measures related to simplicial complexes yields patterns associated to vertices, which themselves contain a set of features. These are used to describe the attacked or the attacker vertices, respectively. Comparing this new strategy with classical concepts demonstrates the advantages of the presented approach using simplicial features for detecting and characterizing intrusions.

Trees Outside Forests (TOF) are an important part of landscapes and agroforestry systems, since they provide a variety of ecosystem services. They contribute to biomass and carbon stocks, support biodiversity conservation and enrich soil. This study investigates the potential of semantic segmentation to distinguish between Forest and the TOF classes Patch, Linear and Tree in aerial images. Information about these classes is crucial in order to understand their distribution and influence on the environment. An automated labeling approach with manual refinement is implemented for generating training and validation data. To map TOFs an adapted Pyramid Attention Network is trained. As this classification task is dependent on contextual information a spatial context postprocessing method is applied. The approach shows reliable results on independent validation data.

In this paper, we deal with the task of change detection for environment maps based on automotive radar sensors. As our main contribution, map misalignment common in the real world due to bad positioning is taken into account as part of the prediction. The data augmentation strategies proposed in this paper imparts the knowledge of rotation and translation invariance into the baseline convolutional neural network based siamese architecture, trained with radar maps of highway scenes focusing on construction zones on the road. Using our proposed strategies, our model outperformed the baseline by attaining 18% and 13% higher average F1 scores with respect to rotational and translational misalignments.

Transfer learning is a standard technique for improving deep learning model performance, in particular when training data is limited. The basic idea is to use the weights learned from an original dataset and transfer these weights to a different (target) dataset. Using an existing state-of-the art method for fault detection, we implemented transfer learning between data sources, conducting extensive experimentation to compare seven fault detection datasets, spanning univariate and multivariate cases. We cross-compared models pre-trained on one dataset and then transferred to other datasets. Our results indicate that, compared to the other datasets, XJTU generalises well in multiple cases when used for pre-training.

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 performance of a deep learning model depends heavily on its architectural hyperparameters. However, there is often little guidance on how to tune those hyperparameters. This paper provides insights into how to tune the architectural hyperparameters of a wide-kernel convolutional model for industrial fault detection, by analysing a grid search over 12,960 possible combinations of hyperparameter settings on seven benchmark datasets of vibration time series. By aggregating the results on these seven datasets, we are able to generalise across multiple industrial fault detection settings. We find that, generally speaking, the number of filters in the later convolutional layers and the hyperparameters associated with the first layer are the most important. Additionally, we analyse the relationships between hyperparameters and develop this analysis into a ‘recommended sequence’ for how to tune them one-at-a-time.

In this paper we present a deep learning based approach for detecting changes or deviations in the context of radar based occupancy grid maps. Specifically, we propose a convolutional neural network (CNN) based architecture to identify spatial changes. As a reference map of the environment, we use occupancy maps generated using detections obtained from automotive radar sensors fitted to the corners of a test-vehicle. For the purpose of similarity learning, a siamese architecture is used. The network is trained with occupancy maps of highway and urban scenes captured over a period of time around the city of Wuppertal, Germany, focusing on construction zones on the road. As per the initial evaluations, the siamese network is able to classify images with construction zones as changes from non-changes i.e. images without construction zones.

The growing number of applications of machine learning and data mining in many domains—from agriculture to business, education, industrial manufacturing, and medicine—gave rise to new requirements for how to inspect and control the learned models. The research domain of explainable artificial intelligence (XAI) has been newly established with a strong focus on methods being applied post-hoc on black-box models. As an alternative, the use of interpretable machine learning methods has been considered—where the learned models are white-box ones. Black-box models can be characterized as representing implicit knowledge—typically resulting from statistical and neural approaches of machine learning, while white-box models are explicit representations of knowledge—typically resulting from rule-learning approaches. In this introduction to the special issue on ‘Explainable and Interpretable Machine Learning and Data Mining’ we propose to bring together both perspectives, pointing out commonalities and discussing possibilities to integrate them.

Athletes are exposed to various psychological and physiological stressors, such as losing matches and high training loads. Understanding and improving the resilience of athletes is therefore crucial to prevent performance decrements and psychological or physical problems. In this review, resilience is conceptualized as a dynamic process of bouncing back to normal functioning following stressors. This process has been of wide interest in psychology, but also in the physiology and sports science literature (e.g. load and recovery). To improve our understanding of the process of resilience, we argue for a collaborative synthesis of knowledge from the domains of psychology, physiology, sports science, and data science. Accordingly, we propose a multidisciplinary, dynamic, and personalized research agenda on resilience. We explain how new technologies and data science applications are important future trends (1) to detect warning signals for resilience losses in (combinations of) psychological and physiological changes, and (2) to provide athletes and their coaches with personalized feedback about athletes' resilience.

In this paper, we apply graph signal processing to optimise a soil moisture sensor network by identifying and re-moving redundant sensors. We evaluated which of seven proposed graph construction techniques best models the relationships between soil moisture measurements at different places in an agricultural field. Here, we consider a sensor location to be redundant if the moisture value can be imputed from information elsewhere in the graph. We gradually remove redundant sensors from the network in a top-down manner, imputing the masked sensors using Tikhonov minimisation - looking for the graph structure that gives us the most accurate imputed values. Our results indicate that the thresholded Gaussian kernel has the best performance in terms of error, while Delaunay triangulation, a parameter-free method, performs similarly. Furthermore, as expected, it seems that the edge sensors are most important while sensors close-by each other or in the centre of the field seem to be less relevant.

Conventionally, event-related potential (ERP) analysis relies on the researcher to identify the sensors and time points where an effect is expected. However, this approach is prone to bias and may limit the ability to detect unexpected effects or to investigate the full range of the electroencephalography (EEG) signal. Data-driven approaches circumvent this limitation, however, the multiple comparison problem and the statistical correction thereof affect both the sensitivity and specificity of the analysis. In this study, we present SHERPA – a novel approach based on explainable artificial intelligence (XAI) designed to provide the researcher with a straightforward and objective method to find relevant latency ranges and electrodes. SHERPA is comprised of a convolutional neural network (CNN) for classifying the conditions of the experiment and SHapley Additive exPlanations (SHAP) as a post hoc explainer to identify the important temporal and spatial features. A classical EEG face perception experiment is employed to validate the approach by comparing it to the established researcher- and data-driven approaches. Likewise, SHERPA identified an occipital cluster close to the temporal coordinates for the N170 effect expected. Most importantly, SHERPA allows quantifying the relevance of an ERP for a psychological mechanism by calculating an ”importance score”. Hence, SHERPA suggests the presence of a negative selection process at the early and later stages of processing. In conclusion, our new method not only offers an analysis approach suitable in situations with limited prior knowledge of the effect in question but also an increased sensitivity capable of distinguishing neural processes with high precision.

Knowledge-augmented learning enables the combination of knowledge-based and data-driven approaches. For anomaly detection and diagnosis, understandability is typically an important factor, especially in high-risk areas. Therefore, explainability and interpretability are also major criteria in such contexts. This chapter focuses on knowledge-augmented explainable and interpretable learning to enhance understandability, transparency and ultimately computational sensemaking. We exemplify different approaches and methods in the domains of anomaly detection and diagnosis - from comparatively simple interpretable methods towards more advanced neuro-symbolic approaches.

We present SD4Py, a free open-source Python package for performing subgroup discovery and analysis. SD4Py makes it easy to discover subgroups from data stored in a Pandas data frame, to undertake follow-on analysis to examine the variability in the quality of the subgroups and to visualise important parameters. The core algorithms for discovering subgroups are implemented by an existing well-established and efficient Java back-end, but are exposed through a user-friendly Python interface. SD4Py offers a concise workflow for not only discovering but also comparing subgroups, in order to select those of interest, and for gaining insights into what is distinctive about individual subgroups.

With their increase in performance, neural network architectures also become more complex, necessitating explainability. Therefore, many new and improved methods are currently emerging, which often generate so-called saliency maps in order to improve interpretability. Those methods are often evaluated by visual expectations, yet this typically leads towards a confirmation bias. Due to a lack of a general metric for explanation quality, non-accessible ground truth data about the model's reasoning and the large amount of involved assumptions, multiple works claim to find flaws in those methods. However, this often leads to unfair comparison metrics. Additionally, the complexity of most datasets (mostly images or text) is often so high, that approximating all possible explanations is not feasible. For those reasons, this paper introduces a test for saliency map evaluation: proposing controlled experiments based on all possible model reasonings over multiple simple logical datasets. Using the contained logical relationships, we aim to understand how different saliency methods treat information in different class discriminative scenarios (e.g. via complementary and redundant information). By introducing multiple new metrics, we analyse propositional logical patterns towards a non-informative attribution score baseline to find deviations of typical expectations. Our results show that saliency methods can encode classification relevant information into the ordering of saliency scores.

Physics Informed Neural Networks (PINNs) feature applications to various partial differential equations (PDEs) in physics and engineering. Many real-world problems contain interfaces, i. e., discontinuities in some model parameter, and have to be included in any relevant PDE solver toolkit. These problems do not necessarily admit smooth solutions. Therefore, interfaces cannot be naturally included into classical PINNs, since their learning algorithm uses the strong formulation of the PDE and does not include solutions in the weak sense. The interface information can be incorporated either by an additional flux condition on the interface or by a variational formulation, thus also allowing weak solutions. This paper proposes new approaches to combine either the weak or energy functional formulation with the piece-wise strong formulation, to be able to tackle interface problems. Our new method PINNTEGRATE can incorporate integrals into the neural network learning algorithm. This novel method cannot only be applied to interface problems but also to other problems that contain an integrand as an optimization objective. We demonstrate PINNTEGRATE on variational minimal surface and interface problems of linear elliptic PDEs.

In general, Graph Neural Networks (GNNs) enable modeling and learning in the context of complex data like graphs and time series – which is typically difficult for standard Deep Learning approaches. This paper proposes an approach for modeling a prediction task using irregularly sampled spatio-temporal sensor data via GNNs. Specifically, we present a method for modeling spatio-temporal sensor data represented as graphs, and a more convenient image representation enabling standard convolutional deep learning. By mapping the irregularly sampled graph to a regular graph representation, we can then integrate temporal sensor information with spatial information. In our experimentation, we demonstrate the efficacy of the proposed approach in inspection contexts of oil and gas pipelines.