Manufacturing decision support requires explanations that not only link predictions to underlying causes but also translate them into feasible actions for stakeholders. Explainable artificial intelligence is therefore pivotal in predictive maintenance, where trust and transparency are essential. However, existing explainable artificial intelligence approaches often emphasize correlations and low-level feature attributions, which can limit both stakeholder trust and practical adoption. In this study, we propose an explainable artificial intelligence framework that bridges the gap between predictive maintenance model outputs and actionable stakeholder-oriented insights. More specifically, we propose a novel framework that automatically extracts causal knowledge graphs from data, enabling explanations grounded in causal reasoning and the suggestion of mitigation actions without human intervention. The framework integrates feature attribution (SHapley Additive exPlanations), temporal causal discovery (VARLiNGAM), and counterfactual analysis to generate intervention-aware explanations and actionable recommendations, supported by large language models. The large language model component constructs semantic taxonomies of signals and translates causal and counterfactual evidence into concise, persona-specific briefs and checklists. To validate the effectiveness of our approach, we evaluate it on two manufacturing use cases using two real-world datasets—AI4I-2020 and Azure-PM—and six models: four tree ensembles (Random Forest, XGBoost, LightGBM, CatBoost), a kernel classifier (Support Vector Machine), and a tabular foundation model (TabPFN). Finally, we evaluate the quality of the large language model-generated explanations and recommendations through a two-tier metrics-based assessment and a state-of-the-art LLM-as-Judge evaluation protocol.

Although recent generative models can produce time series with close marginal distributions, they often face a fundamental tension between preserving global temporal structure and modeling stochastic local variations, particularly for highly volatile signals with weak or irregular periodicity. Direct distribution matching in such settings can amplify noise or suppress meaningful temporal patterns. In this work, we propose a structure-residual perspective on time-series generation, viewing temporal data as the combination of a structural backbone and stochastic residual dynamics, thereby motivating the separation of global organization from sample-level variability. Based on this insight, we represent time-series structure using a quantile-based transition graph that compactly captures global distributional and temporal dependencies. Building on this representation, we propose Graph2TS, a quantile-graph conditioned variational autoencoder that performs cross-modal generation from structural graphs to time series. By conditioning generation on structure rather than labels or metadata, the model preserves global temporal organization while enabling controlled stochastic variation. Experiments on diverse datasets, including sunspot, electricity load, ECG, and EEG signals, demonstrate improved distributional fidelity, temporal alignment, and representativeness compared to diffusion- and GAN-based baselines, highlighting structure-controlled and cross-modal generation as a promising direction for time-series modeling.

Today, high-performance computing (HPC) systems play a crucial role in advancing artificial intelligence. Nevertheless, the estimated global data center electricity consumption in 2022 was around 1% of the final global electricity demand. Therefore, as HPC systems advance towards Exascale computing, research is required to ensure their growth is sustainable and environmentally friendly. Data from infrastructure monitoring can be leveraged to predict downtimes, ensure these are treated in time, and increase the overall system's utilization. In this paper, we compare four machine-learning approaches, three of them based on graph embeddings, to predict compute node downtimes. The experiments were performed with data from Marconi 100, a tier-0 production supercomputer at CINECA in Bologna, Italy. Our results show that the machine learning models can accurately predict downtime, matching current state-of-the-art models.

In many cases, a machine learning model must learn to correctly predict a few data points with particular values of interest in a broader range of data where many target values are zero. Zero-inflated data can be found in diverse scenarios, such as lumpy and intermittent demands, power consumption for home appliances being turned on and off, impurities measurement in distillation processes, and even airport shuttle demand prediction. The presence of zeroes affects the models' learning and may result in poor performance. Furthermore, zeroes also distort the metrics used to compute the model's prediction quality. This paper showcases two real-world use cases (home appliances classification and airport shuttle demand prediction) where a hierarchical model applied in the context of zero-inflated data leads to excellent results. In particular, for home appliances classification, the weighted average of Precision, Recall, F1, and AUC ROC was increased by 27%, 34%, 49%, and 27%, respectively. Furthermore, it is estimated that the proposed approach is also four times more energy efficient than the SOTA approach against which it was compared to. Two-fold models performed best in all cases when predicting airport shuttle demand, and the difference against other models has been proven to be statistically significant.

Data is vital in enabling machine learning models to advance research and practical applications in finance, where accurate and robust models are essential for investment and trading decision-making. However, real-world data is limited despite its quantity, quality, and variety. The data shortage of various financial assets directly hinders the performance of machine learning models designed to trade and invest in these assets. Generative methods can mitigate this shortage. In this paper, we introduce a set of novel techniques for time series data generation (we name them Fiaingen) and assess their performance across three criteria: (a) overlap of real-world and synthetic data on a reduced dimensionality space, (b) performance on downstream machine learning tasks, and (c) runtime performance. Our experiments demonstrate that the methods achieve state-of-the-art performance across the three criteria listed above. Synthetic data generated with Fiaingen methods more closely mirrors the original time series data while keeping data generation time close to seconds - ensuring the scalability of the proposed approach. Furthermore, models trained on it achieve performance close to those trained with real-world data.

While graph sampling is key to scalable processing, little research has tried to thoroughly compare and understand how it preserves features such as degree, clustering, and distances dependent on the graph size and structural properties. This research evaluates twelve widely adopted sampling algorithms across synthetic and real datasets to assess their qualities in three metrics: degree, clustering coefficient (CC), and hop plots. We find the random jump algorithm to be an appropriate choice regarding degree and hop-plot metrics and the random node for CC metric. In addition, we interpret the algorithms' sample quality by conducting correlation analysis with diverse graph properties. We discover eigenvector centrality and path-related features as essential features for these algorithms' degree quality estimation, node numbers (or the size of the largest connected component) as informative features for CC quality estimation and degree entropy, edge betweenness and path-related features as meaningful features for hop-plot metric. Furthermore, with increasing graph size, most sampling algorithms produce better-quality samples under degree and hop-plot metrics.

Visual Quality Inspection is an integral part of the manufacturing process that is becoming increasingly automated with the advent of Industry 4.0. While very beneficial, AI-driven Computer Vision Algorithms and Deep Neural Networks face several issues that may impede their adoption in practical real-life settings such as a manufacturing shop floor. One such issue arising during an AI classifier’s continuous operation is the frequent lack of robustness to novel defects appearing for the first time. Such unanticipated inputs can pose a significant risk to cyber-physical applications as a resulting out-of-context decision could compromise the integrity of the production process. While recent Machine Learning methods can theoretically tackle this problem from different angles (e.g., open-set recognition, semi-supervised learning, intelligent data augmentation), applying them to a real-life setting with a small, imbalanced dataset and high inter-class similarity can be challenging. This paper confronts such a use case aiming at the automation of the visual quality inspection of shaver shell brand prints from the electronics industry and characterized by data scarcity and the existence of small local defects. To that end, we introduce a novel data augmentation approach based on the latent space manipulation of StyleGAN, where defect data is intentionally synthesized to simulate novel inputs that can help form a boundary of the model’s knowledge. Our approach shows promising results compared to well-established open-set recognition and semi-supervised methods applied to the same problem, while its consistent performance across classifier embeddings indicates lower coupling to the final classifier.

The increasing size of graph structures in real-world applications, such as distributed computing networks, social media, or bioinformatics, requires appropriate sampling algorithms that simplify them while preserving key properties. Unfortunately, predicting the outcome of graph sampling algorithms is challenging due to their irregular complexity and randomized properties. Therefore, it is essential to identify appropriate graph features and apply suitable models capable of estimating their sampling outcomes. In this paper, we compare three machine learning (ML) models for predicting the divergence of five metrics produced by twelve node, edge, and traversal-based graph sampling algorithms: degree distribution (D3), clustering coefficient distribution (C2D2), hop-plots distribution (HPD2) (including the largest connected component (HPD2C)), and execution time. We use these prediction models to recommend suitable sampling algorithms for each metric and conduct mutual information analysis to extract relevant graph features. Experiments on six large real-world graphs demonstrate a prediction error under 20 % in C2D2 and HPD2 prediction for most algorithms despite their relatively high dissimilarity with the training data. Sampling algorithm recommendations on ten real-world graphs show higher hits@3 for D3 and C2D2 and comparable results for HPD2 and HPD2C compared to the K-best baseline method. Finally, ML models show superior runtime recommendations compared to baseline methods, with hits@3 over 86 % for synthetic and real graphs and hits@ 1 over 60 % for small graphs. These findings are promising for algorithm recommendation systems, particularly when balancing quality and runtime preferences.

Manufacturing industries strive to improve production efficiency and product quality by deploying advanced sensing and control systems. Wearable sensors are emerging as a promising solution for achieving this goal, as they can provide continuous and unobtrusive monitoring of workers' activities in the manufacturing line. This paper presents a novel wearable sensing prototype that combines IMU and body capacitance sensing modules to recognize worker activities in the manufacturing line. To handle these multimodal sensor data, we propose and compare early, and late sensor data fusion approaches for multi-channel time-series convolutional neural networks and deep convolutional LSTM. We evaluate the proposed hardware and neural network model by collecting and annotating sensor data using the proposed sensing prototype and Apple Watches in the testbed of the manufacturing line. Experimental results demonstrate that our proposed methods achieve superior performance compared to the baseline methods, indicating the potential of the proposed approach for real-world applications in manufacturing industries. Furthermore, the proposed sensing prototype with a body capacitive sensor and feature fusion method improves by 6.35%, yielding a 9.38% higher macro F1 score than the proposed sensing prototype without a body capacitive sensor and Apple Watch data, respectively.

Quality control is a key activity performed by manufacturing enterprises to ensure products meet quality standards and avoid potential damage to the brand’s reputation. The decreased cost of sensors and connectivity enabled an increasing digitalization of manufacturing. In addition, artificial intelligence enables higher degrees of automation, reducing overall costs and time required for defect inspection. In this research, we compare three active learning approaches and five machine learning algorithms applied to visual defect inspection with real-world data provided by Philips Consumer Lifestyle BV. Our results show that active learning reduces the data labeling effort without detriment to the models’ performance.

Explainable Artificial Intelligence promises to deliver means so that humans better understand the rationale behind a particular machine learning model. In the image domain, such information is frequently conveyed through heat maps. Along the same line, information regarding defect detection for unsupervised methods applied to images can be conveyed through anomaly maps. Nevertheless, heat maps or anomaly maps can convey inaccurate information (artifacts), or their perceptions may differ across different persons. Therefore, the user experience could be enhanced by collecting human feedback and creating predictive models on how these could be recolored to bridge the gap between the original heat maps and anomaly maps created with explainability techniques and the output expected by humans. We envision this work as relevant in at least two scenarios. First, enhance anomaly and heat maps when conveying information regarding machine vision models deployed in production to remove information deemed unnecessary by the user but systematically present through the explainability technique due to underlying model issues (artifacts). Second, adapt anomaly and heat maps based on users’ perceptual needs and preferences.

Quality control is being increasingly automatised in the context of Industry 4.0. Its automatisation reduces inspection times and ensures the same criteria are used to evaluate all products. One of the challenges when developing supervised machine learning models is the availability of labelled data. Few-shot learning promises to be able to learn from few samples and, therefore, reduce the labelling effort. In this work, we combine this approach with unsupervised methods that learn anomaly maps on unlabelled data, providing additional information to the model and enhancing the classification models' discriminative capability. Our results show that the few-shot learning models achieve competitive results compared to those trained in a classical supervised classification setting. Furthermore, we develop novel active learning data sampling strategies to label an initial support set. The results show that using sampling strategies to create and label the initial support set yields better results than selecting samples at random. We performed the experiments on four datasets considering real-world data provided by Philips Consumer Lifestyle BV and Iber-Oleff - Componentes Tecnicos Em Plastico, S.A.

As high-performance computing (HPC) systems advance towards Exascale computing, their size and complexity increase, introducing new maintenance challenges. Modern HPC systems feature data monitoring infrastructures that provide insights into the system’s state. This data can be leveraged to train machine learning models to anticipate anomalies that require compute nodes to undergo maintenance procedures. This paper presents a novel approach to predicting such anomalies by creating a graph per measurement that encodes current and past sensor readings and information related to the compute node sensors. The experiments were performed with data collected from Marconi 100, a tier-0 production supercomputer at CINECA in Bologna, Italy. Our results show that the machine learning model can accurately predict anomalies and surpass current State-Of-The-Art (SOTA) models regarding the quality of predictions and the time horizon considered to forecast them.

Visual defect recognition and its manufacturing applications have been an upcoming topic in recent AI research. Defect datasets are often severely imbalanced and can be additionally burdened with separating classes of high visual similarity. Although various methods of data augmentation have been proposed to mitigate the class imbalance, they often fail to cope with tinier minority classes or have fidelity issues with smaller defects while, at the same time, needing significant computational resources to train. Also, augmentation based on vector-based oversampling struggles to produce high-fidelity inputs and is hard to apply on custom CNN architectures, which often perform better for this type of problem. Our work presents an image-level oversampling method based on an instance-based image generator that can be applied to any CNN directly during the training process without increasing the order of training time required. It is based on identifying a small number of the most uncertain base samples close to the estimated class boundaries and using them as seeds for augmentation. The resulting images are of high visual quality preserving small class differences, and they also improve the classifier boundary leading to higher recall scores than other state-of-the-art approaches.

Manufacturing workers' fatigue is an acknowledged concern with implications for well-being, health, safety, and operational performance. Past studies have employed physiological measurements obtained from smartwatches and wearable devices, seeking to assess and classify the fatigue state of workers. However, the extent to which models developed based on data obtained from individual workers could apply to other workers remains unclear. This paper presents the results of an exploratory study in which data from different subjects are employed to develop a range of fatigue estimation and predictive machine learning models. A cross-subject study provides evidence of sufficiently accurate performance in several cases. Further insights arise from looking into cases of lower generalization and linking these to personal characteristics. Copyright (c) 2024 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)