Högskolan i Skövde

Many real-world machine learning classification problems suffer from imbalanced training data, where the least frequent label has high relevance and significance for the end user, such as equipment breakdowns or various types of process anomalies. This imbalance can negatively impact the learning algorithm and lead to misclassification of minority labels, resulting in erroneous actions and potentially high unexpected costs. Most previous oversampling methods rely only on the minority samples, often ignoring their overall density and distribution in relation to the other classes. In addition, most of them lack in the oversampling method’s explainability. In contrast, this paper proposes a novel oversampling method that considers a subspace of the feature-set for the creation of synthetic minority samples using nonlinear optimization of a class-sensitive objective function. Suitable subspaces for oversampling are identified through a Bayesian reinforcement strategy based on Dirichlet smoothing, which may be useful for explainable-AI. An empirical comparison of the proposed method is performed with 10 existing techniques on 18 real-world datasets using two traditional machine learning classifiers and four evaluation metrics. Statistical analysis of cross-validated runs over the 18 datasets and four metrics (i.e. 72 experiments) reveals that the proposed approach is among the best performing methods in 6 and 2 instances when using random forest classifier and support vector machine classifier, thus placing it at the top. The study also reveals that some feature combinations are more important than others for minority oversampling, and the proposed approach offers a way to identify such features.

This paper presents a hybrid machine learning framework to enhance traffic safety in urban road intersections. The framework employs a two-stage approach: Decision Trees predict vehicle trajectories by identifying turning behaviors, while Random Forests estimate collision probabilities involving vehicles and vulnerable road users (VRUs) such as pedestrians and cyclists. Engineered spatial, temporal, and motion-related features are derived from high-resolution trajectory data collected via connected camera systems in busy urban cores. The experimental results demonstrate high predictive accuracy, achieving an F1-Score of 0.97 for turning vehicle classification and a ROC-AUC of 0.98 for collision risk estimation. Compared to computationally intensive deep learning models, the proposed framework balances robust performance with computational efficiency, making it suitable for realtime deployment in complex urban environments. The framework integrates with in-vehicle Human-Machine Interfaces (HMIs) to enhance driver awareness and enable proactive safety interventions. This study addresses the need for interpretable and scalable road safety solutions in connected traffic systems.

Driver intention recognition (DIR) studies increasingly rely on deep neural networks. Deep neural networks have achieved top performance for many different tasks. However, apart from image classifications and semantic segmentation for mobile phones, it is not a common practice for components of advanced driver assistance systems to explicitly analyze the complexity and performance of the network’s architecture. Therefore, this paper applies neural architecture search to investigate the effects of the deep neural network architecture on a real-world safety critical application with limited computational capabilities. We explore a pre-defined search space for three deep neural network layer types that are capable to handle sequential data (a long-short term memory, temporal convolution, and a time-series transformer layer), and the influence of different data fusion strategies on the driver intention recognition performance. A set of eight search strategies are evaluated for two driver intention recognition datasets. For the two datasets, we observed that there is no search strategy clearly sampling better deep neural network architectures. However, performing an architecture search improves the model performance compared to the original manually designed networks. Furthermore, we observe no relation between increased model complexity and better driver intention recognition performance. The result indicate that multiple architectures can yield similar performance, regardless of the deep neural network layer type or fusion strategy. However, the optimal complexity, layer type and fusion remain unknown upfront.

Machine tools are essential to manufacturing for precise and efficient component production. With Industry 4.0, abundant machine condition data enables data-driven maintenance decisions. However, deploying condition-based maintenance solutions is challenging due to the diverse configurations of equipment, complex failure modes, and compatibility issues with the digital infrastructure. While machine tool health monitoring relies on detailed tests like Ballbar measurements, they consume valuable production time. To address these challenges, this article presents a human-centric development and deployment of a condition-based data-driven maintenance dashboard. The solution uses data from the controller system to improve machine tool testing in a Swedish heavy-duty vehicle powertrain facility.

Predictive maintenance is a central concept in the shift towards Industry 4.0. Accurately estimating the remaining useful life of a machine, or a machine component, is an important aspect of predictive maintenance. Deep learning models have previously been applied to this task with success. However, these models may not perform well for cases where training data is sparse. In these situations, the model should also provide some degree of uncertainty about its prediction to instill trust in the user. Hence, predictive models should accurately estimate their own uncertainty, in addition to providing correct predictions. In this paper, we propose up-sampling of sparse ballbar measurement data in order to generate adequate samples to train and evaluate deep neural networks. The inference is conducted with three different types of models, Monte Carlo Dropout, variational inference, and deep ensemble. The approaches are compared based on point prediction accuracy, and uncertainty quantification quality. It is found that both Monte Carlo Dropout and deep ensemble perform well in regards to predictive accuracy, with the deep ensemble consistently resulting in the best calibrated uncertainty estimation. 

The health of a machine tool directly affects its ability to produce components with high precision. Therefore, monitoring and diagnosing early faults can enhance its reliability resulting in an improvement in manufacturing throughput and overall product quality. This paper concerns condition monitoring of the ballscrew drive, a machine tool component that transforms rotary motion of the drive shaft into linear motion of the work table along the guideways. The degradation of the ballscrew drive is often characterized by backlash, which results in imprecise linear motion and, therefore, affects the position of guideways during machining operations. Many physical characteristics of the ballscrew drive, such as required torque, viscous friction, and Coulomb friction, change with the degradation of the ballscrew during its lifetime. The paper proposes a condition monitoring methodology consisting of four main steps: data collection, data preprocessing and feature engineering, model building, and anomaly detection. The machine tool drive system is operated under no-load condition at regular intervals to capture health data using Siemens Analyze MyCondition instrumentation. Subsequently, the data is preprocessed and features are extracted from raw signals using the wavelet transform approach. The unsupervised machine learning technique, principal component analysis, is used to reduce the dimensionality of the dataset and find feature combinations that capture most of the variation in the data. Next, Hotelling’s T2 statistic is computed for each sample on a rolling basis, and anomalous behavior is detected based consistent deviations beyond the moving median of Hotelling’s T2 statistic. The proposed methodology is applied on condition monitoring data from a Swedish automotive manufacturer and the health assessments are validated against backlash measurements obtained from a different conditional monitoring test. This shows that the health status of a ballscrew can be derived directly from its physical characteristics.

Driver intention recognition (DIR) methods mostly rely on deep neural networks (DNNs). To use DNNs in asafety-critical real-world environment it is essential to quantify how confident the model is about the producedpredictions. Therefore, this study evaluates the performance and calibration of a temporal convolutionalnetwork (TCN) for multiple probabilistic deep learning (PDL) methods (Bayes-by-Backprop, Monte-Carlodropout, Deep ensembles, Stochastic Weight averaging - Gaussian, Multi SWA-G, cyclic Stochastic GradientHamiltonian Monte Carlo). Notably, we formalize an approach that combines optimization-based pre-trainingwith Hamiltonian Monte-Carlo (PT-HMC) sampling, aiming to leverage the strengths of both techniques. Ouranalysis, conducted on two pre-processed open-source DIR datasets, reveals that PT-HMC not only matchesbut occasionally surpasses the performance of existing PDL methods. One of the remaining challenges thatprohibits the integration of a PDL-based DIR system into an actual car is the computational requirements toperform inference. Therefore, future work could focus on optimizing PDL methods to be more computationallyefficient without sacrificing performance or the ability to estimate uncertainties.

Climate change has impacted global temperatures, triggering extreme weather and adverse environmental effects. In Sweden, these changes have caused shifts in weather patterns, leading to disruptions in infrastructure. This, in turn, has influenced water turbidity levels, negatively impacting water quality. To tackle these issues, a study was conducted using machine learning to predict turbidity with six meteorological variables collected for two years. Our preliminary research showed a substantial influence of seasonal changes on water turbidity, especially air temperature. Identifying supporting indicators such as lagged features is crucial and considerably improved the turbidity prediction performance for two of the machine learning models used. However, the study also identified challenges like data collection and uncertainty issues. We recommend improving data collection quality with higher frequency, minimizing geographical gaps between data collection points, sharing calibration assumptions, checking the sensors regularly, and accounting for data anomalies. Understanding these challenges and their potential implications could lead to more methodological enhancements.

Dataanalys och det närbesläktade området artificiell intelligens utvecklas snabbt och är av stort intresse för ledningsområdet då de kan erbjuda åtråvärda egenskaper som automatiserad hantering av stora datamängder och modeller för att tolka ny data. Den här rapporten omfattar en avskanning av fyra utvalda delområden: preskriptiv analys, osäkerhetshantering, förklarbarhet (XAI) och systemperspektiv. Rapporten innehåller dessutom en uppskattning av den framtida utvecklingen av delområdena och en jämförelse med avseende på deras mognadsgrad och relevans.

Data analysis and the closely related field of artificial intelligence are developing rapidly and are of great interest to the command and control field as they can offer desirable features such as automated handling of large data sets and models for interpreting new data. This report includes a scan of four selected sub-areas: prescriptive analysis, uncertainty management, explainability (XAI) and systems perspective. The report also contains an estimate of the future development of the sub-areas and a comparison with respect to their degree of maturity and relevance.

Real-world applications of artificial intelligence that can potentially harm human beings should be able to express uncertainty about the made predictions. Probabilistic deep learning (DL) methods (e.g., variational inference [VI], VI last layer [VI-LL], Monte-Carlo [MC] dropout, stochastic weight averaging - Gaussian [SWA-G], and deep ensembles) can produce a predictive uncertainty but require expensive MC sampling techniques. Therefore, we evaluated if the probabilistic DL methods are uncertain when making incorrect predictions for an open-source driver intention recognition dataset and if a surrogate DL model can reproduce the uncertainty estimates. We found that all probabilistic DL methods are significantly more uncertain when making incorrect predictions at test time, but there are still instances where the models are very certain but completely incorrect. The surrogate DL models trained on the MC dropout and VI uncertainty estimates were capable of reproducing a significantly higher uncertainty estimate when making incorrect predictions.