Högskolan i Skövde

Complex manufacturing processes generate high-dimensional data, necessitating continuous monitoring, analysis, and diagnosis of quality issues to ensure improved product quality. Data-driven quality operations have gained significant attention with recent advances in digitalization. In this context, supervised machine learning methods can be applied due to the availability of labeled data. However, quality data are often imbalanced because defect classes are typically underrepresented, and is called class imbalance. To overcome this, oversampling methods generate synthetic minority samples to balance the class distribution, thereby improving the performance of classification algorithms on unseen data. However, synthetic samples generated in the interior regions of the minority class may not help in improving classification performance. Rather, synthetic samples generated near the boundaries of these regions may lead to better classifiers. To test this hypothesis, this paper proposes a technique for generating synthetic minority samples near class decision boundaries. Our approach first detects minority samples near class boundaries in various subspaces of the feature-set and uses them as initial points for optimizing a class-sensitive objective function designed to generate additional minority samples along those boundaries. The probabilities of selecting feature subspaces are updated through Bayesian reinforcement during oversampling, and important features are identified as those forming the most reinforced subspaces. The proposed technique is compared with state-of-the-art oversampling methods on 15 imbalanced benchmark datasets and validated on a real-world foundry dataset. Experimental results show that it achieves the best overall performance with Random Forest and Support Vector Machine classifiers. On the foundry dataset, it attains the highest F₁ score, demonstrating its effectiveness.

Predictive maintenance (PdM) plays an important role in the transition toward Industry 4.0, offering the potential to increase operational efficiency and reduce unexpected downtime. A central application of PdM involves the estimation of a system component’s remaining useful life (RUL) to enable informed maintenance decisions. Deep learning (DL) models have demonstrated strong performance in RUL prediction tasks. However, these models often lack the capability to reliably quantify predictive uncertainty, which is essential for risk-aware decision-making in industrial applications. Bayesian approximation models, like Stochastic Weight Averaging Bayesian and Variational Inference, can yield a distribution of predictions while avoiding the prohibitively costly process of true Bayesian inference. This distribution can be used to reason about the uncertainty of the model, but often shows the model being overconfident. We propose empirical likelihood stacking, a method for improving uncertainty quantification for DL models. By applying stacking and deep ensembles to estimate the RUL of turbofan engines, we show the trade-offs between these two methods when applied to DL models. We find that the stacked model has a better calibrated uncertainty at a small cost to point prediction accuracy, when compared to the ensemble.

Manufacturing companies are gradually moving from Industry 4.0’s technology focus to Industry 5.0’s sustainability focus, and identifying and addressing energy bottlenecks is a part of this transition. In practice, this is challenging due to limited availability of energy data and its poor integration with systems like MES and SCADA. Energy dashboards are capable of consolidating energy data, visualizing consumption patterns, and tracking related KPIs for sustainability. However, most existing implementations are limited to facility-level overviews or machine-specific views without consideration of operational details. To identify energy bottlenecks, the dashboards must also analyze machine states, batch sizes, product mixes, and cycle times. Therefore, this paper presents a Python-based web application built with the Dash framework and open-source packages. The application integrates data from EMS, MES, and SCADA systems. It is capable of performing statistical time-series analysis, joint energy-stop analysis, state-based mapping of energy use, and visualizing various Key Performance Indicators. The proposed integrated dashboard targets discrete manufacturing and is demonstrated on a gear machining line at Volvo Group Trucks Operations. The dashboard currently operates offline with data from enterprise systems, but aims for real-time API integration as a digital twin in the future. This could support simulation for detecting inefficiencies, predicting energy bottlenecks, and optimizing energy consumption.

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.

Human–robot collaboration solutions have gradually become popular, in which some tasks are performed by robots and others by humans. Designing such a production cell requires simultaneous consideration of human-centered factors, machine-focused mechanical design, and system engineering in the early planning stages. However, different objectives often conflict (e.g., speeding up a robot can improve productivity while compromising energy efficiency), and the same variables can affect multiple models and simulations simultaneously (e.g., a machine where humans and robots collaborate can influence both the operator’s working posture and the robot’s cycle time). Therefore, multidisciplinary tools and multi-level optimization are needed to model, simulate, and optimize elements such as production flows, robotics, and human operators to balance objectives related to cycle time, energy consumption, and worker well-being. In this paper, we formulate an approach that integrates different simulation tools and a bi-level optimization framework to balance worker well-being, cycle time, and energy consumption. We demonstrate this approach through a real industrial case of designing a work cell for elevator pipe assembly in a grain conveying system, where ABB RobotStudio is used for robotic simulation and IPS IMMA for human simulation. IBM ILOG CPLEX Optimization Studio is employed for the top-level task allocation optimization, and a set of results is presented based on data extracted from the lower-level robot-centered optimization. The results show that our approach can effectively balance different objectives by incorporating detailed information from different levels of the work cell design.

Factory layout planning is central to manufacturing performance. By aligning productivity with worker well-being, factories can achieve more sustainable operations. Traditional approaches often treat the process sequence as static, limiting adaptability. This article introduces a human-centric approach that treats the process sequence as reconfigurable, enabling dynamic adjustment of resources and equipment across the assembly line. The approach integrates sequence generation with precedence constraints, block layout creation and resource positioning into a unified multi-objective optimization process. A genetic algorithm combined with digital human modelling evaluates proposals in terms of worker well-being, walking distance and area utilization, while ensuring compliance with work-environment regulations. Unlike previous studies, this method integrates reconfigurable sequencing with resource positioning into full assembly-line optimization, validated in an industrial testbed. The findings show how adaptable layouts improve performance and embed human factors and ergonomics into decision making, supporting Industry 5.0 principles of sustainable, human-centric manufacturing.

Ergonomics evaluation methods are crucial for assessing risks for work-related musculoskeletal disorders and ensuring operator well-being, productivity, and safety. Despite the increased use of digital twins and AI-supported tools in production system design and operation, ergonomics evaluations still primarily rely on observational techniques such as expert assessments or checklist-based tools like the Rapid Entire Body Assessment and the Rapid Upper Limb Assessment. These methods are time-consuming, imprecise, and prone to subjectivity arising from variability in the judgment of the ergonomists and ambiguity in scoring criteria. As an alternative, ergonomics evaluation methods based on using technologies for direct measurements can provide semi-automation of the assessments and offer greater objectivity and precision. This study investigates the capability of a computer vision-based motion capture approach to support direct measurement ergonomics evaluations and compares its results with those of an inertial measurement unit-based system in an industrial task. The comparison was conducted by studying output data of the two systems and by feeding the data into a direct measurement-based ergonomics evaluation method. A representative industrial assembly task involving upper-body movement and dynamic wrist activity was recorded simultaneously using a single monocular RGB camera and an IMU-based system. Both datasets were processed using two parallel workflows that followed the same structure to extract joint angles and segment positions over time. The comparison and evaluation of the results demonstrates that computer vision-based motion capture has the potential to provide human posture and motion data suitable for direct measurement ergonomics evaluations in industrial environments.

Accurate anomaly detection and localization in sheet metal glue line applications are crucial for quality assurance in automotive manufacturing. Most current vision-based inspection systems that rely on geometric deviations from a predefined shape often suffer from high false-positive rates, leading to unnecessary interventions and operational inefficiencies. This research investigates the potential of unsupervised deep learning models to significantly reduce false positives in the analysis of sheet metal glue line images, even with limited datasets. We conducted a comparative evaluation of 17 unsupervised deep learning models covering different categories with 28 backbones on datasets of approximately 300 industrial glue line images per part from a Swedish vehicle manufacturer. A data synthesis method was applied to balance the glue line dataset, further enhancing the reliability of the models. To address the challenge of limited training data and improve model generalization, we incorporated data augmentation techniques and performed robustness experiments to ensure applicability to real-world industrial conditions. Our findings demonstrate that deep learning approaches can effectively detect and localize anomalies, significantly reducing false positives and gluing machine downtimes compared to the existing system. Moreover, we proposed a multi-criteria decision-making based approach for model selection, enabling decision-makers to achieve optimal trade-offs between accuracy and inference time, thus improving operational efficiency. These advancements highlight that even with limited training data, unsupervised deep learning models can enhance anomaly detection reliability, streamline the automotive production process, and reduce unnecessary resource expenditures.

Virtual manufacturing, simulation, and optimization provide a wealth of knowledge about the possibilities of future production systems so as to support decision makers. However, this knowledge usually remains with a handful of domain experts, is not captured and is hard to share even within the same team. At the same time, simulations can benefit from the incorporation of linked data from real factories once a process is running. Graph databases provide a possible approach to storing and managing this form of interrelated heterogeneous data, with powerful querying capabilities that can identify important or interesting patterns that might otherwise remain hidden. Current research focuses on one or two aspects of this problem but does not address all at once, despite the potential benefits of the combination. This paper provides a broad literature review of the current directions within research with a special focus on how graphs can support finding knowledge within Virtual Factories, used by larger teams for industrial planning and optimization.

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.