Automatic fault detection and diagnosis (FDD) are critical for maintaining reliable and efficient industrial systems. However, conventional methods rely heavily on manual inspections or threshold-based techniques, which often fail to capture the dynamic patterns in time series (TS) sensor data. As a result, faults persist for extended periods, leading to suboptimal system operations, increased energy waste, and significant economic losses. This work proposes a cross-modal framework that facilitates the efficient deployment of state-of-the-art pretrained vision models for enhanced FDD, with two novel TS-to-image transformations: first, an adapter deep encoder that learns optimal, task-specific representations from raw sensor data while generating outputs that are input-compliant with pretrained models. Second, an enhanced line plot that creates geometric shapes of two related signals. Comparative experiments against fixed methods, including spectrograms, Gramian angular fields, Markov transition fields, recurrence plots, and five deep learning baseline models, showed substantial performance gains across diverse domains. InceptionTime achieved the highest average baseline performance with an F<inf>1</inf> of 88.6%, while the adapter and shapes achieved 94.4% and 92.4%, respectively. The findings highlight the potential of the cross-modal framework for FDD to facilitate early intervention and efficient system maintenance in industrial settings.

Fault detection in district heating (DH) substations is critical for energy efficiency and reliability. However, it is challenged by scarce fault labels, low-frequency data, privacy concerns, and battery-constrained gateways. We propose a novel hybrid semi-supervised federated domain adaptation architecture for fault detection in DH. We use a one-class variational autoencoder (VAE) to leverage heterogeneous sensor streams from 434 distributed substations. First, we perform cross-network unsupervised pre-training on multi-sourced data from two independent real-world DH networks, fusing their return temperature dynamics into a robust shared manifold. Second, we leverage maintenance metadata to selectively allow verified-normal clients for per-round fine-tuning of the model. Third, we drastically reduce uplink costs by compressing each client's weight delta using 10% top-k sparsification and demonstrate that our pipeline enables robust few-shot finetuning with 20% of the normal operational data while retaining high detection performance. By strategically training, our method achieves F1 and G-mean scores of up to 97% and an AUC ≥ 99% on real-world DH data. To our knowledge, this is the first work to study cross-domain data fusion in the DH field for fault detection, aiming to enhance and enable effective, scalable, and energy-efficient monitoring of substations.

This paper proposes SCENTS, a Streaming Consensus Embedding for Time Series that fuses streaming windows into a single low-dimensional representation suitable for diverse downstream tasks. First, it learns a denoised low-dimensional latent basis for state initialisation (). Second, for each newly arriving stream, it performs a near-linear time, multi-pass consensus update that fuses new affinities directly into Z. We prove convergence and validate our assumptions using various real-world multi-source industrial datasets on a streaming consensus clustering task. In contrast to conventional pipelines that accumulate noise over time, require costly  co-association matrices, or  eigendecompositions, SCENTS yields a linear memory consensus embedding that improves monotonically along the streaming horizon and produces high-quality partitions. Moreover, SCENTS is designed to be integrated with any fixed encoder, enabling the implementation of a lightweight streaming consensus adaptation layer. The resulting embeddings offer a compact, reusable representation that supports various downstream tasks beyond clustering, thereby providing a scalable and generalisable fusion layer for data analysis.

Keyword spotting (KWS) in historical documents is an important tool for the initial exploration of digitized collections. Nowadays, the most efficient KWS methods rely on machine learning techniques, which typically require a large amount of annotated training data. However, in the case of historical manuscripts, there is a lack of annotated corpora for training. To handle the data scarcity issue, we investigate the merits of self-supervised learning to extract useful representations of the input data without relying on human annotations and then use these representations in the downstream task. We propose ST-KeyS, a masked auto-encoder model based on vision transformers where the pretraining stage is based on the mask-and-predict paradigm without the need for labeled data. In the fine-tuning stage, the pre-trained encoder is integrated into a fine-tuned Siamese neural network model to improve feature embedding from the input images. We further improve the image representation using pyramidal histogram of characters (PHOC) embedding to create and exploit an intermediate representation of images based on text attributes. The proposed approach outperforms state-of-the-art methods trained on the same datasets in an exhaustive experimental evaluation of five widely used benchmark datasets (Botany, Alvermann Konzilsprotokolle, George Washington, Esposalles, and RIMES). 

The authentication of scanned administrative documents is a critical challenge in digital forensics due to the accessibility of sophisticated image editing tools. While existing forgery detection methods focus on geometric patterns or hardware signatures, they often struggle with the structural complexity of Arabic cursive script or require original reference templates. This paper proposes a novel framework based on statistical intensity continuity analysis for detecting forgeries in scanned Arabic documents. The method targets the physical distinction between the stochastic noise introduced by analog scanning and the artificial uniformity of digital manipulation. We introduce a novel maximum continuous intensity value (MCIV) feature and utilize interquartile range (IQR) analysis to identify statistical outliers in text block intensity distributions. The proposed pipeline was evaluated on a newly created dataset of 609 Arabic administrative scans acquired through a multiscanner strategy. Experimental results using 10-fold cross-validation demonstrate that the random forest classifier achieves a peak accuracy of 84.70% and a forged-class F1-score of 89%. Comparative benchmarking against established methods from the literature shows that the proposed framework achieves a superior precision of 89.31% and significantly reduces the false positive rate (FPR) to 24.82%, providing a more reliable and defensible balance for forensic applications. Furthermore, the system demonstrates high computational efficiency (0.197 s per document), making it suitable for high-volume forensic triage in real-world administrative environments.

Large amount of data are generated from in-situ monitoring of additive manufacturing (AM) processes which is later used in prediction modelling for defect classification to speed up quality inspection of products. A high volume of this process data is defect-free (majority class) and a lower volume of this data has defects (minority class) which result in the class-imbalance issue. Using imbalanced datasets, classifiers often provide sub-optimal classification results i.e. better performance on the majority class than the minority class. However, it is important for process engineers that models classify defects more accurately than the class with no defects since this is crucial for quality inspection. Hence, we address the class-imbalance issue in manufacturing process data to support in-situ quality control of additive manufactured components.  For this, we propose cluster-based adaptive data augmentation (CADA) for oversampling to address the class-imbalance problem. Quantitative experiments are conducted to evaluate the performance of the proposed method and to compare with other selected oversampling methods using AM datasets from an aerospace industry and a publicly available casting manufacturing dataset. The results show that CADA outperformed random oversampling and the SMOTE method and is similar to random data augmentation and cluster-based oversampling. Furthermore, the results of the statistical significance test show that there is a significant difference between the studied methods.  As such, the CADA method can be considered as an alternative method for oversampling to improve the performance of models on the minority class. 

This paper presents a conceptual design for a virtual reality (VR)-based furniture arrangement system, proposed to leverage immersive technologies and generative AI (GenAI) with intuitive interaction methods. The system utilizes a head-mounted display (HMD) to provide an immersive and intuitive user experience (UX) with multimodal interaction methods through hand and eye-tracking controls. Additionally, GenAl enables the virtual agent to engage in natural conversations with users and interpret and respond contextually, aiming to enhance personalization by understanding conversations. Users' interests and preferences are analyzed and predicted from the conversation and eye-gaze data, which provides recommendations for relevant furniture items and real-time personalized feedback. By combining these techniques, this design aims to create a seamless, interactive, intelligent, personalized virtual interior design and furniture arrangement experience within the immersive virtual environment (VE). The implementation of several key features demonstrates a proof of concept for our virtual furniture arrangement system. 

Diabetic retinopathy (DR) is a leading cause of blindness worldwide, necessitating early detection to prevent severe visual impairment. Despite numerous proposed classification techniques, challenges persist due to the high parameter count of deep learning algorithms, imbalanced datasets, and limited performance. This study introduces a novel framework for DR classification that leverages multi-view deep features, multilinear whitened principal component analysis, tensor exponential discriminant analysis, synthetic minority oversampling technique, and deep random forest. We evaluated this architecture using the APTOS blindness dataset under a standard protocol. The results demonstrate that our architecture significantly improves classification accuracy, surpassing existing methods. Our contributions highlight a promising approach for enhancing DR classification performance.

Personality trait identification through handwriting analysis presents a challenging area within automated document recognition based on Artificial Intelligence solutions. Recent studies relied on solutions automating graphonomic processes, while others address only a few local features, conversely few studies offer solutions based on textural features. In this work, we propose an automated approach for personality trait identification that treats a scripter’s handwriting as a texture by leveraging a diverse set of textural features, including LCP, oBIFCs, LPQ, LBP, among others. The approach is validated on FFM-annotated datasets using cost-effective classifiers such as XGBoost, Random Forest, Gradient Boost, SVM, and Naive Bayes. Our empirical study enabled the judicious selection of the most suitable textural features for each personality trait. Subsequently, we constructed a comprehensive personality trait identification solution by combining multiple textural features and integrating top-performing classifiers. The experimental results demonstrated the validity of our hypothesis, achieving performance improvements of more than 10% on both datasets. 

Background: In addition to impairing the lung function, chronic obstructive pulmonary disease (COPD) also affects phonatory characteristics. Recent research highlights the potential of voice as a digital biomarker to support clinical decision-making. While machine learning (ML) can detect disease patterns from acoustic features, clinical relevance requires understanding the relationship between the disorder and acoustic features.

Objective: This study investigates both statistical and clinical significance using Baseline Acoustic (BLA) and Mel-Frequency Cepstral Coefficient (MFCC) features with focusing on individuals with COPD and healthy controls (HC).

Method: Acoustic features derived from Swedish utterances of the vowel [a:], recorded via mobile phones from 48 age-matched participants (24 COPD, 24 HC; equal gender distribution), were analyzed. To reduce bias from varying recording counts, features were aggregated by averaging 10 randomly selected recordings per participant over 100 iterations. Vowel articulation was visualized in the vowel quadrilateral space using F1 (tongue height) and F2 (tongue advancement). Group differences were assessed using the Shapiro-Wilk test, Mann-Whitney U test (α = 0.05), Benjamini-Hochberg (BH) and Bonferroni corrections, Permutational Multivariate Analysis of Variance (PERMANOVA) test, and Cliff's Delta (δ).

Results: Of 101 features, 29 remained significant after BH correction and one after Bonferroni. Multivariate testing (p = 0.019) showed group separation. Additionally, 34 features demonstrated large effect sizes, suggesting potential as digital biomarkers.

Conclusion: Voice data recorded via mobile phones capture meaningful acoustic differences associated with COPD. These findings support the integration of voice-based assessments into eHealth platforms for noninvasive COPD screening and monitoring, which is pending further validation on larger populations.

Clinical Trial: NCT06705647