https://www.mdu.se/

Machine learning is an increasingly popular method for modeling complex systems. A common machine learning model is the neural network, which can be trained to represent complicated functions to a high accuracy. However, neural networks often grow large and complex. Recent work is looking at how to abstract networks to yield simpler representations while retaining some property of the original network — for instance, such that for every input the abstracted network's output is at least as large as the original. In this work, we build on previous ideas and extend them to also consider the input layer. Sometimes the input vector has a large size while only a few of the elements are significant in the computation of the output. We propose to use a trained neural network model to identify insignificant input elements, i.e., elements which do not contain important information. We show how the presented abstraction method for the input layer can be utilized to achieve this.

Effective water management in underground mining requires maintaining safe reservoir levels while minimizing the high energy costs of continuous pumping. Although flexible electricity pricing enables cost-aware operation, traditional threshold-based controllers cannot exploit this flexibility efficiently. This paper presents an industrial case study on efficient mine dewatering using reinforcement-learning-based control synthesized with the Uppaal Stratego framework. A baseline threshold controller is first implemented, followed by a reinforcement-learning controller trained on forecast inflows and day-ahead electricity prices to minimize pumping costs while limiting pump switching. To ensure safety during learning without distorting the optimization objective, we introduce a pre-shield that blocks unsafe transitions. We formally show that this pre-shield is maximally permissive with respect to a monotonicity safety objective. Simulation results demonstrate that the learning-based strategy reduces total energy consumption by up to 40% compared to threshold-based control, while maintaining safe operation in all scenarios. © The Author(s) 2026

Artificial Intelligence (AI) is widely used in Industry 4.0 and Industry 5.0 applications for anomaly detection and predictive analytics. However, the opaque decision-making of most Machine Learning (ML) models limits interpretability and hinders effective decision-making and root cause analysis. Explainable AI (XAI) seeks to address these challenges; however, the combined processes of anomaly detection and explanation generation often require substantial domain expertise, limiting their accessibility to a broader audience. To address this gap, this paper presents EXACT (EXplainable Anomaly Classification Tool), a novel, modular software framework that enables practitioners and researchers to rapidly and easily perform explainable anomaly detection on time-series data, regardless of domain expertise. EXACT integrates time-series anomaly detection with XAI techniques into a single end-to-end workflow, including data handling, anomaly injection, parameter tuning, model training, anomaly detection, explanation generation, and visualization, making these features broadly accessible and user-friendly to the wider community. The effectiveness of EXACT is demonstrated through rigorous evaluations conducted on three public time-series datasets from distinct domains, using three anomaly detection models and three widely adopted XAI methods. Among these, XGBoost exhibits the most balanced predictive performance, achieving 99.96% accuracy, 82.61% F1-score, and 98.59% ROC-AUC on the Credit Card Fraud dataset, while Decision Trees demonstrate significantly lower training times. In terms of explainability, SHAP consistently delivers high-quality feature rankings, reaching an NDCG score of 0.98 on the Credit Card Fraud dataset, whereas LIME provides substantially faster explanation generation. These results demonstrate that EXACT effectively facilitates explainable anomaly detection in an efficient, user-friendly manner, while also providing a systematic analysis of trade-offs among predictive performance, interpretability, and computational efficiency. EXACT is publicly available at GitHub (https://github.com/TedBoman/EXACT), and the community is invited to contribute to its continued development and extension.

The validation of Automated Driving Systems (ADSs) has shifted from distance-based metrics to Scenario-Based Testing (SBT). Large Language Models (LLMs) have emerged as powerful tools with potential for generating vehicular scenarios at scale. However, generative models used for direct simulation synthesis produce inadequate output, therefore necessitating a more structured compilation approach. We present HASCO (Hybrid AI Simulation COmpiler), a system that translates natural-language driving scene specifications into executable simulation artifacts (XOSC/XODR files) for the esmini/OpenSCENARIO ecosystem. While LLMs excel at narrative parsing, we demonstrate that direct synthesis of simulation artifacts fails in the vast majority of cases due to hallucinated physics or schema violations. To resolve this, HASCO treats scenario creation as a compilation task rather than a generative one. The pipeline supports three compilation paths: direct synthesis, a Python intermediate (via scenariogeneration), and an ontology-guided path that grounds intent into an intermediate representation before compilation. We further evaluate a self-judging mechanism for automated repair. Across six operating modes evaluated on 40 real-world accident reports, the ontology-guided and Python-based compilers achieve 95% and 90% executability rates, respectively, compared to 5% for direct synthesis. We additionally evaluate outputs on semantic fidelity, positioning HASCO as a robust tool for forensic scene reconstruction.

Atrial Fibrillation (AFib) is a common cardiac arrhythmia whose global prevalence has risen in recent years. If left untreated, AFib can lead to severe complications such as stroke and heart failure. Because AFib may occur without symptoms, continuous monitoring is critical for timely detection. This paper presents a low-cost Intelto detect AFib from Electrocardiogram (ECG) signals. The study evaluates the feasibility of deploying 1D-CNN models on resource-constrained edge devices and compares edge- and cloud-based computing architectures with respect to inference efficiency. Three 1D-CNN models of increasing complexity are designed, trained, and tested on datasets containing AFib and Normal Sinus Rhythm (NSR) segments. Two experiments are conducted to assess end-to-end delay and prediction time under a controlled experimental setup. The results demonstrate the potential for on-device AFib detection in constrained environments and provide practical insights into selecting suitable architectures for embedded deployment.

To address safety concerns in industrial systems, we propose a framework for forecasting CPU load with respect to a predetermined threshold, allowing customers to add tasks from a predefined library. Existing tools, akin to Windows Task Manager, provide limited insights due to their aggregate nature and high computational overhead. Our approach uses conformal prediction for rapid uncertainty-aware forecasts and Shapley value analysis to quantify individual task contributions to the CPU load. This proof-of-concept framework improves system safety assessment by addressing key research questions in load prediction and validation, paving the way for refined measurement methodologies in industrial applications.

This paper introduces and discusses, at a conceptual level, a smart self-adaptation framework for distributed computing systems, with self-adaptive and assurance capabilities, which uses industrial-scale digital twins as a validation and analytics platform. One of the relevant innovations introduced by our work consists in the combo adoption of Artificial Intelligence, formal verification and domain-specific languages to meaningfully magnify the overall performance, accuracy and resilience of the framework. The paper concludes with open research issues and future research directions of the investigated research area.

Engineering Digital Twins (EDT) presents a multifaceted challenge that extends beyond managing the lifecycle of a Digital Twin (DT) to include its continuous, dynamic interaction with the lifecycle of the actual object, system, or process it represents, referred to as the Actual Twin (AT). The relationship between the lifecycles of DT and AT necessitates a rethinking of the software development lifecycle of DTs. This vision paper examines the deeply intertwined lifecycles of DT and AT, arguing that effective methods for EDT must embrace the mutual and adaptive evolution of both over time. We propose placing evolution at the core of EDT. We identify key triggers of DT evolution, examine the engineering dimensions involved, and explore how the best practices, technologies, and tools of DevOps can support this evolution. Finally, we discuss current challenges and opportunities in the field. This paper serves as a call to action for the EDT community to adopt evolution as a crucial factor and core principle in EDT.

Given increased complex data with high dimensions, feature selection aims to select a subset of features to increase the efficiency of machine learning. This paper proposes a new feature selection method based on membrane computing. The proposed method has two main advantages. First, it provides a new solution to search for feature combinations while requiring no model construction (which is time-consuming) to evaluate a feature subset. Second, feature selection is embedded in a membrane clustering algorithm, which is designed to enable searching for the best feature subset and finding active cluster centres at the same time. The designed clustering algorithm mimics the behavior of multiple cells and it has stronger global search ability than existing evolutionary algorithms. The efficacy of the proposed method has been shown by the evaluation of a set of benchmark data sets.

Integrating machine learning into computer architecture simulation offers a new approach to performance analysis, moving away from traditional algorithmic methods. While existing simulators accurately replicate hardware, they often suffer from slow execution, complex documentation, and require deep CPU knowledge, limiting their usability for quick insights. This paper presents a deep learning-based approach for simulating a key CPU component, cache memory. Our model "learns" cache characteristics by observing cache miss distributions, without needing detailed manual modeling. This method accelerates simulations and adapts to different program needs, demonstrating accuracy comparable to traditional simulators. Tested on Sysbench and image processing algorithms, it shows promise for faster, scalable, and hardware-independent simulations.