https://www.mdu.se/

Artificial Intelligence (AI) is reshaping decision-making across industries, but in life-critical domains like Air Traffic Control (ATC), performance alone is not enough - AI must also be transparent, accountable, and human-centered. At the same time, ecological sustainability (e.g., fuel efficiency, emissions reduction, noise abatement) is becoming a key operational priority. Yet under high workload or time pressure, these goals are often deprioritized. This paper introduces a novel framework for bias-aware human-AI teaming in ATC, integrating Generative Explainable AI (genXAI), Predictive Telemetry (PT), multimodal machine learning, and mechanistic interpretability. Designed for Air Traffic Controllers (ATCOs), the system anticipates future traffic states, detects bias-prone conditions, and delivers adaptive, context-sensitive explanations. Through bias-aware interfaces, serious-game training, and neuroadaptive feedback loops, the framework aligns decision support with cognitive state and operational complexity. The framework targets three key objectives: (1) mitigating cognitive biases, (2) improving ground routing decisions, and (3) validating anticipatory AI models. By reframing sustainability as a human-AI collaboration challenge, this work advances a new class of trustworthy, bias-aware AI that enhances both safety and ecological performance in Air Traffic Management (ATM). 

Learning over time for machine learning (ML) models is emerging as a new field, often called continual learning or lifelong Machine learning (LML). Today, deep learning and neural networks are the prevalent approaches for LML models. However, they are often criticized for being “black box” methods, and catastrophic forgetting has remained a persistent challenge throughout their development. This paradigm represents a significant shift from traditional static learning models, enabling systems to adapt to new data continuously while retaining previously acquired knowledge. In this paper, an LML approach is proposed that combines a Random Forest (RF) with a Genetic Algorithm (GA) to transfer the knowledge from an existing RF model to a new learning model. Here, the GA is applied to the RF model so that the weights of this model get balanced more steadily. The approach is evaluated here for classification problems on three benchmark datasets. The initial results present knowledge retention in the new model, indicating the success of the model. These methods are regularization-based and effective in mitigating catastrophic forgetting. However, they rely on Fisher Information to estimate parameter importance or similar measures, which are computationally demanding and suitable for deep neural networks. While many NN-based lifelong learning approaches have been studied, RFs remain comparatively underexplored in this paradigm. Given their robustness, interpretability, and effectiveness in tabular datasets with limited samples, RFs present a compelling alternative. This motivates our work on developing an RF-based lifelong learning approach. 

Industrial environments generate large volumes of heterogeneous, unstructured PDF documents, such as material certificates, making reliable automated data extraction a methodological challenge. Traditional LLM-based pipelines often struggle with noisy OCR output, inconsistent layouts, and schema variability, leading to unstable outputs and reduced accuracy. This study addresses the problem of designing a robust, consistent extraction pipeline capable of operating across diverse document structures under both limited and scaled domain-specific supervision regimes. We propose a coordinated multi-agent framework integrating layout-aware OCR processing, expert-guided prompt engineering, systematic prompt refinement, controlled synthetic dataset expansion, and parameter-efficient adaptation via LoRA. The architecture includes specialised agents for raw text extraction, dataset preparation, prompt optimisation, LoRA configuration search, output processing, and structured evaluation, forming an iterative workflow that improves stability and reduces error propagation. Structural compliance and retrieval accuracy are evaluated separately using binomial modelling with Wilson confidence intervals. Zero-shot, few-shot, and LoRA fine-tuned LLMs are assessed across commercial and open-weight models. Results show that few-shot prompting enables rapid structural alignment under low-resource conditions, while fine-tuning effectiveness depends on supervision diversity. Merged supervision, defined as fine-tuning on combined original and synthetic certificates (FT-Merged), substantially outperforms limited-scale LoRA adaptation. Under FT-Merged, Mistral 7B achieves 97.05% retrieval accuracy with near-perfect structural compliance, approaching commercial reliability. GPT-4o and DeepSeek Chat achieve approximately 99% accuracy without task-specific fine-tuning. Cross-domain validation confirms structural generalisation, and runtime-cost analysis highlights trade-offs between accuracy, latency, financial cost, and deployment control. Overall, robust industrial extraction emerges from coordinated architectural orchestration rather than model choice alone. © 2026

Explainability is crucial for establishing user trust in Artificial Intelligence (AI), particularly within safety-critical domains such as Air Traffic Management (ATM) and Air Traffic Control (ATC). This study empirically investigates the effects of Explainable AI (XAI), specifically HeatMap-based visual explanations, on cognitive workload, user acceptance, and intention to use AI-driven decision-support systems among Air Traffic Control Officers (ATCOs). Despite significant theoretical advancements in the broader XAI domain, empirical evidence addressing the specific impact of visual explanations on human-AI interactions in safety-critical environments like ATC remains limited. To address these critical gaps, an experimental comparison was conducted between explainable (HeatMap) and non-explainable (BlackBox) AI conditions, involving two user groups: expert and student ATCOs. Both objective neurophysiological measures (Electroencephalography) and subjective questionnaires were employed to capture comprehensive user responses. Key findings revealed that the presence of visual explanations significantly reduced cognitive workload and enhanced users' willingness to adopt the AI system, regardless of participants' level of expertise. However, explicit perceptions of AI's impact on work performance were predominantly influenced by expertise, with less experienced controllers reporting a greater perceived impact than their expert counterparts. By combining objective neurometrics with subjective user assessments, this research advances methodological rigor in evaluating human-AI interactions and highlights the importance of tailored, user-centric explanations. These findings directly contribute to practical guidelines for designing cognitively compatible and trustworthy AI tools in ATC, providing nuanced insights for targeted training and deployment strategies based on user expertise. 

Fault detection and diagnosis in complex industrial systems, such as slurry pumps used in mining, is impractical with lack of labeled data. However, most of the existing studies rely on labelled data and the black-box nature of machine learning or deep learning approaches. While some studies have focused on model-specific explanations, domain-specific explanations are still missing, which hinders actionable maintenance decisions. To address those challenges, this study focuses on an explainable multi-stage self-supervised learning framework for intelligent fault discovery and diagnosis with three sequential phases: (1) a self-supervised module that generates high-confidence labels by fusing vibration sensor and process data through autoencoders, Isolation Forest, and Gaussian Mixture Models with cross-modality consensus, (2) a supervised classification phase enhanced with SHAP-based explainability to distinguish between healthy and anomalous operation with high accuracy, and (3) a fault confirmation and diagnosis layer that categorizes severity (Tolerable, Moderate, Severe) and identifies likely fault types using domain knowledge and shallow decision trees. An experimental study evaluated on real-world slurry pump data demonstrates that the framework successfully generated 19,369 healthy and 654 anomalous training labeled samples. The supervised models achieved exceptional performance, with 99.9% accuracy and perfect recall (100%) in distinguishing healthy from anomalous states. The severity categorization identified 58.6% cases as tolerable (potentially edge-case operations), 27.8% as severe, and 13.6% as moderate, while fault type identification revealed cavitation or impeller damage as the most prevalent fault (46%). This approach bridges data-driven learning and expert knowledge, enabling trustworthy, explainable, and actionable fault discovery and diagnosis in safety-critical industrial environments.

Lithium-ion battery state estimation and prognostics are crucial for the safe and reliable operation of electric vehicles (EVs), renewable energy systems, and portable devices. However, nonlinear battery behavior complicates State of Charge (SOC) estimation, limited and imbalanced data hinder State of Health (SOH), and scarce degradation trajectories with domain shift challenge Remaining Useful Life (RUL) prediction. Generative Adversarial Networks (GANs) offer an effective approach by generating realistic synthetic data that mitigates scarcity, improves diversity, and enhances model robustness. This study aims to systematically review and consolidate the current literature on GAN-based data augmentation for battery state estimation and prognostics. Specifically, it examines the effectiveness of different GAN architectures and techniques in improving SOC, SOH, and RUL estimation, assesses synthetic data quality and reliability, identifies technical challenges and limitations, and outlines evidence-based guidelines and future research directions. A systematic literature review was conducted following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, incorporating relevant studies obtained from four prominent digital libraries. Analysis of 31 primary studies reveals that widely used public datasets (National Aeronautics and Space Administration (NASA), Center for Advanced Life Cycle Engineering (CALCE) and Oxford) dominate the field, enabling reproducibility and benchmarking. GAN-based augmentation achieves 17% to 90% error reductions across root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and mean squared error (MSE) metrics. Time-series GANs and Wasserstein GANs (WGANs) emerge as most effective, with Adam optimizer learning rate (LR) = 0.001 and gradient penalty (λ<inf>GP</inf>=10) providing improved stability. Major challenges include data scarcity, synthetic data quality concerns, GAN instability, and domain shift. Six priority areas are identified for advancing the field: physics-informed constraints, domain adaptation, uncertainty quantification, real-time deployment, multimodal learning, and data efficiency. This review establishes GAN-based augmentation as significant for battery state estimation and prognostics. It provides evidence-based insights and a roadmap towards developing reliable, interpretable, and deployable battery management systems. © 2026 Elsevier Ltd

Predictive maintenance is critical for minimizing downtime and operational costs in mining industries, where slurry pumps operate under abrasive and highly variable conditions. Traditional machine learning models, while accurate, often lack interpretability, limiting their adoption in safety-critical environments. This paper presents an Adaptive Neuro-Fuzzy Inference System (ANFIS) based approach for anomaly detection in slurry pumps using vibration-based features such as displacement, velocity, crest factor, kurtosis, skewness, peak, and peak-to-peak values. ANFIS combines the transparency of fuzzy logic with the learning capability of neural networks, enabling interpretable IF-THEN rules and adaptive tuning of membership functions. A dataset of 33,104 vibration samples was analyzed under four balancing strategies to address class imbalance. Models were evaluated using accuracy, precision, recall, F1-score, RMSE, and clustering quality metrics (Dunn Index and Silhouette Score). The results demonstrate that ANFIS achieves high accuracy (>99%) and strong interpretability, outperforming traditional black-box models. The proposed approach enhances trustworthiness and adaptability in predictive maintenance systems, offering an explainable solution for Industry 4.0 applications. Future work will explore hybrid models, real-time IoT integration, and edge deployment for dynamic operational environments.

Rectified linear unit (ReLU) based neural networks (NNs) are recognised for their remarkable accuracy. However, the decision-making processes of these networks are often complex and difficult to understand. This complexity can lead to challenges in error identification, establishing trust, and conducting thorough analyses. Existing methods often fail to provide clear insights into the actual computations occurring within each layer of these networks. To address this challenge, this study introduces a mechanistic interpretability method called ReLU Region Reasoning (Re3). This method uses the known piecewise-linear characteristics of ReLU networks to offer insights into neuron activation and accurately assess how each feature contributes to the final output and probability. Re3 effectively determines neuron activations and evaluates the contribution of each feature within a specified linear region. Experiments conducted on multiple benchmark datasets, including both tabular and image data, demonstrate that Re3 can replicate individual predictions without error, align feature importance with domain expertise, and maintain consistency with current explanatory methods, thereby avoiding the typical randomness. Analysing neurons reveals activation sparsity and identifies dominant units, thus providing clear targets for model simplification and troubleshooting. By ensuring transparency and algebraic accessibility in each stage of a ReLU-based NN's decision process, Re3 can be a valuable practical tool for achieving precise mechanistic interpretability.

Transformer feed-forward (MLP) blocks perform nonlinear transformations, but their internal structure is often difficult to interpret using current methods. This paper presents anexact soft-gate decomposition that extracts the effective weight matrix for each MLP blockat varying inputs. This approach generalises exact piecewise-affine decomposition fromReLU networks to Gaussian Error Linear Unit (GELU) and Sigmoid Linear Unit (SiLU)activations, covering both standard and gated MLP architectures. Verification across fivemodels, including BERT-base, mBART-50, and GPT-2 Small for language tasks, ViT-basefor vision tasks, and the gated SiLU architecture TinyLlama-1.1B, confirms the accuracyof this method at floating-point arithmetic precision. The application of singular value decomposition to the effective weight matrix yields six structural metrics that characteriseeach block based on the dimensionality, intensity, and spectral concentration of its transformation. When applied to one model from each of three architectural configurations,BERT-base, mBART-50, and GPT-2 Small, alongside ablation as causal ground truth, thisanalysis indicates that MLP computation is organised differently across the models studied.Specifically, BERT-base’s blocks vary by transformation dimensionality, mBART-50’s blocksby transformation intensity, and GPT-2 Small’s blocks by intensity with negative ablationvalues in the later layers. These structural differences, revealed only through the exact effective weight matrix, explain why scalar importance measures often fail to generalise acrossthe models studied.