Juan Galán-Páez

Few-Shot Learning (FSL) has gained significant attention in fault diagnosis due to its ability to classify faults with limited labeled data. Metric learning-based approaches, such as prototypical networks, have demonstrated effectiveness in this domain by computing class prototypes from support examples. However, these methods often rely on high-dimensional embeddings, which can lead to computational inefficiencies and overfitting, particularly in low-data regimes. This work introduces the Regularized Embedding Memory Book (REMB), a novel module designed to improve prototype estimation and enhance classification accuracy in FSL for fault diagnosis. The REMB module constructs a memory book of embeddings for each class and refines prototype computation by selecting the most representative embeddings. Additionally, we incorporate a set of regularization techniques to improve generalization and robustness, particularly in noisy environments. We evaluate REMB on four benchmark datasets commonly used in fault diagnosis: CWRU, JNU, PU, and an additional dataset with varying training sample sizes. The experimental results demonstrate that REMB improves classification performance, particularly in low-data scenarios, and enhances model calibration. Furthermore, we analyze the impact of noise adaptation and the contribution of different regularization terms through an ablation study. Our findings suggest that the proposed approach provides a robust and efficient solution for few-shot fault diagnosis, outperforming conventional metric learning methods.

Predictive maintenance, within the field of Prognostics and Health Management (PHM), aims to identify and anticipate potential issues in equipment before they become serious problems. Deep Learning (DL) models, such as Deep Convolutional Neural Networks (DCNN), Long Short-Term Memory (LSTM) networks, and Transformers, have been widely adopted for this task and have shown great success. However, these models are often considered "black boxes" due to their opaque decision-making processes, making it challenging to explain their outputs to industrial equipment experts. The complexity and vast number of parameters in these models further complicate understanding their predictions. This paper introduces a novel Explainable AI (XAI) algorithm, an extension of the well-known Local Interpretable Model-agnostic Explanations (LIME). Our approach uses a conditioned Probabilistic Diffusion Model to generate altered samples in the neighborhood of the original sample studied. We validate our method using various rotating machinery diagnosis datasets. Additionally, we compare our approach with state-of-the-art XAI methods, employing nine metrics to evaluate the desirable properties of any XAI method.

Background/Objectives: The aim of this study was to design a fully automated hybrid AI-based method, combining a convolutional neural network (CNN) and a tree-based model (XGBoost), which was capable of diagnosing different pelvic organ prolapses (POPs) in a dynamic two-dimensional ultrasound study from the midsagittal plane. Methods: This was a prospective observational study with 188 patients (99 with POP and 89 without POP). Transperineal pelvic floor ultrasound videos were performed, and normality or POP was defined. These videos were subsequently labeled, and an algorithm was designed to detect POP based on three phases: 1. Segmentation—a CNN was used to locate and identify the visible pelvic organs in each frame of the ultrasound video. The output had a very high dimensionality. 2. Feature engineering and dataset construction—new features related to the position and shape of the organs detected using the CNN were generated. 3. The POP predictive model—this was created from the dataset generated in the feature engineering phase. To evaluate diagnostic performance, accuracy, precision, recall, and F1-score were considered, along with the degree of agreement with the expert examiner. Results: The best agreements were observed in the diagnosis of cystocele and uterine prolapse (88.1%) and enterocoele (81.4%). The proposed methodology showed an accuracy of 96.43%, an overall accuracy of 98.31%, a recall of 100%, and an F1-score of 98.18% in detecting the presence of POP. However, when differentiating between the various types of POP, we observed that the precision, accuracy, recall, and F1-score were higher when detecting cystocele and uterine prolapse. Conclusions: We have developed the first predictive model capable of diagnosing POP in a dynamic, bi-dimensional ultrasound study from the midsagittal plane using deep learning and machine learning techniques.

This paper introduces a novel framework for generating synthetic time-series data from turbine engine sensor readings using a text-conditioned diffusion model. The approach begins with dataset preprocessing, including correlation analysis, feature selection, and normalization. Principal Component Analysis (PCA) transforms the normalized signals into three components, mapped to the RGB channels of an image. These components, combined with engine identifiers and cycle information, form compact 19 × 19 × 3 pixel images, later scaled to 512 × 512 × 3 pixels. A variational autoencoder (VAE)-based diffusion model, fine-tuned on these images, leverages text prompts describing engine characteristics to generate high-quality synthetic samples. A reverse transformation pipeline reconstructs synthetic images back into time-series signals, preserving the original engine-specific attributes while removing padding artifacts. The quality of the synthetic data is assessed by training Remaining Useful Life (RUL) estimation models and comparing performance across original, synthetic, and combined datasets. Results demonstrate that synthetic data can be beneficial for model training, particularly in the early epochs when working with limited datasets. Compared to existing approaches, which rely on generative adversarial networks (GANs) or deterministic transformations, the proposed framework offers enhanced data fidelity and adaptability. This study highlights the potential of text-conditioned diffusion models for augmenting time-series datasets in industrial Prognostics and Health Management (PHM) applications.

Purpose: To create and validate a machine learning(ML) model that allows for identifying the correct capture of the midsagittal plane in a dynamic ultrasound study, as well as establishing its concordance with a senior explorer and a junior explorer.Methods: Observational and prospective study with 90 patients without pelvic floor pathology. Each patient was given an ultrasound video where the midsagittal plane of the pelvic floor was recorded at rest and during the Valsalva maneuver. A segmentation model was used that was trained on a previously published article, generating the segmentations of the 90 new videos to create the model. The algorithm selected to build the model in this project was XGBoost(Gradient Boosting). To obtain a tabular dataset on which to train the model, feature engineering was carried out on the raw segmentation data. The concordance of the model, of a junior examiner and a senior examiner, with the expert examiner was studied using the kappa index.Results: The first 60 videos were used to train the model and the last 30 videos were reserved for the test set. The model presented a kappa index 0.930(p < 0.001) with very good agreement for detection of the correct midsagittal plane. The junior explorer presented a very good agreement (kappa index = 0.930(p < 0.001)). The senior explorer presented a kappa index 0.789(p < 0.001) (good agreement) for detection of the correct midsagittal plane.Conclusion: We have developed a model that allows determining the correct midsagittal plane captured through dynamic transperineal ultrasound with a level of agreement comparable to or greater than that of a junior or senior examiner, using expert examiner assessment as the gold standard.

A persistent challenge in machine learning is the computational inefficiency of neural architecture search (NAS), particularly in resource-constrained domains like predictive maintenance. This work introduces a novel learning-curve estimation framework that reduces NAS computational costs by over 50% while maintaining model performance, addressing a critical bottleneck in automated machine learning design. By developing a data-driven estimator trained on 62 different predictive maintenance datasets, we demonstrate a generalized approach to early-stopping trials during neural network optimization. Our methodology not only reduces computational resources but also provides a transferable technique for efficient neural network architecture exploration across complex industrial monitoring tasks. The proposed approach achieves a remarkable balance between computational efficiency and model performance, with only a 2% performance degradation, showcasing a significant advancement in automated neural architecture optimization strategies.

Maintaining equipment and machinery in industries is imperative for maximizing operational efficiency and prolonging their lifespan. The adoption of predictive maintenance enhances resource allocation, productivity, and product quality by proactively identifying and addressing potential equipment anomalies through rigorous data analysis before they escalate into critical issues. Consequently, these measures strengthen market competitiveness and generate favorable economic outcomes. In many applications, sensors operate at high frequencies or capture data over extended periods. This work introduces CONELPABO (Composite Networks Learning via Parallel Bayesian Optimization), a framework for analyzing long time series data, particularly for predicting the remaining useful life of a system or component. It uses a divide-and-conquer strategy to manage the exponential growth in the hyperparameter search space during Bayesian Optimization and to accelerate model training by 50%. Additionally, this strategy enables the training of deeper networks with limited resources. The usefulness of the framework is demonstrated through two case studies, in which it achieves state-of-the-art results, showing that CNN-CNN and RNN-RNN architectures are highly effective for long time-series data. These architectures outperform many existing approaches and challenge the common academic focus on CNN-RNN hybrids.

Este capítulo presenta el proceso de innovación docente aplicado en la asignatura Informática del Grado en Estadística, centrado en la visualización de datos con R. Se siguió una metodología basada en proyectos sobre el mercado inmobiliario de Sevilla.

It has been demonstrated that the amount of data is crucial in data-driven machine learning methods. Data is always valuable, but in some tasks, it is almost like gold. This occurs in engineering areas where data is scarce or very expensive to obtain, such as predictive maintenance, where faults are rare. In this context, a mechanism to generate synthetic data can be very useful. While in fields such as Computer Vision or Natural Language Processing synthetic data generation has been extensively explored with promising results, in other domains such as time series it has received less attention. Previous works have proposed techniques like geometric transformations, interpolation, and generative models like GANs and VAEs for time series data augmentation. However, the use of denoising models, particularly Diffusion Probabilistic Models (DPMs), remains largely unexplored in this context. This work specifically focuses on studying and analyzing the use of different techniques for data augmentation in time series for classification and regression problems. The proposed D3A-TS methodology involves the use of DPMs, which have recently achieved successful results in the field of Image Processing, for data augmentation in time series. Additionally, the use of meta-attributes to condition the data augmentation process is investigated. The results highlight the high utility of this methodology in creating synthetic data to train classification and regression models. To assess the results, six different datasets from diverse domains were employed, showcasing versatility in terms of input size and output types. Finally, an extensive ablation study is conducted to further support the obtained outcomes.