Fernando Sancho-Caparrini

Cyclostationary signals define the operational dynamics of rotating machinery, yet their periodic–stochastic structure and the scarcity of fault labels hinder robust representation learning for practical condition monitoring. We introduce the Signal-World-Model, a self-supervised framework that offers a practical solution to vibration analysis as a latent-state transition problem. The architecture integrates a convolutional encoder with a non-overlapping receptive-field strategy and a predictive transformer to infer the underlying system dynamics without explicit signal reconstruction. Pretrained via a teacher-student-predictor paradigm, the model is adapted for fault diagnosis using minimal labeled data through a lightweight classifier. Experimental validation on a multi-stage centrifugal pump and the Case Western Reserve University bearing case study demonstrates consistent improvements over representative contrastive-based, reconstruction-based, and standard joint-embedding predictive architecture approaches, providing an effective pathway for machinery monitoring under severe data constraints.

Anomaly detection task is an open-set challenge, aiming to identify unseen faulty signals using only healthy signals for training. While data reconstruction frameworks are inherently suited for this task, they often struggle with complex signals due to limited feature extraction capabilities. Contrastive learning offers powerful representation learning, but faces challenges in one-class scenarios and requires effective augmentation techniques. To address these limitations, a novel fault detector is proposed, integrating a redesigned one-class contrastive loss into a data reconstruction framework to endow clustering capability. Learnable feature augmentation is incorporated to ensure effective and generalizable contrastive learning while reducing computational costs by performing augmentation in the feature space. A dilated inception network structure is employed to capture long-distance dependencies in complex input signals. A one-class similarity distance-based threshold is introduced to filter outliers in the healthy signal distribution, and an optimal model selection strategy is proposed based on the minimal threshold during training. The approach is evaluated using two case studies: single-fault and multifault scenarios in a reciprocating compressor. Our method achieves balanced accuracies of 98.44% and 97.81%, respectively, outperforming other methods. These results confirm our detector's capacity to balance false alarms and missed detections effectively, even in challenging multifault conditions.

Selection is often considered as a fundamental force in the evolutionary process. Genetic drift, or selection noise, is an important characteristic of selection methods. It has a direct effect on the performance of genetic algorithms. In this paper, a brief review of methods to analyze genetic drift is given, and known estimations of selection noise of various selection schemes used in genetic algorithms are presented. After that, genetic drift of widely used proportional, ranking, and tournament selection schemes is thoroughly studied. To this end, two new measures for selection noise analysis are proposed, namely the noise takeover time and pure reproduction rate. Using these measures, the effect of population size, chromosome length, and selection scheme parameters on genetic drift is analyzed. Also, selection schemes known as being selection pressure equivalent are tested for selection noise equivalence. Both theoretical and experimental approaches are used for the analysis. The results obtained are presented in tabular form. Wherever possible, it is indicated whether the obtained results are identical or different from the results of previous studies. Since no comprehensive study of selection noise has been conducted previously, this indication concerns only some of the results. Although our results differ at some points from those presented earlier, they are consistent on both measures.

Although neural-symbolic techniques for symbol grounding in sensory data have shown significant effectiveness, they still require substantial training. This article revisits symbolic-only approaches by introducing a new algorithm for creating hierarchical concept structures from spatial sensory data. The method is based on Bateson’s idea of difference as the fundamental element of concept formation. By leveraging this principle, the algorithm extracts atomic features from raw data through basic sequential comparisons within a stream of multivariate numerical values. Experiments carried out on a set of common objects indicate that the method can successfully discriminate and assimilate categories as needed without training. The results show that the model improves on the neural networks it has been tested against, which required more than 400 training examples for the same task. The results also show that the model can generate rich conceptual structures and human-like representations, which (i) facilitate high composability, (ii) support formal reasoning, (iii) inherently enable generalisation, and (iv) possess potential for generative behaviour. Consequently, this approach offers a compelling contribution to symbol grounding and neural-symbolic research by providing a seamless algorithm to bridge perception and conceptual knowledge.

Data-driven fault detection in prognostic and health management research poses significant challenges, particularly with few measurements to model complex machinery health conditions. Approaches requiring data in one or many faulty conditions are impractical for actual use cases where only healthy condition data exist. Most of the existing methods for this scenario rely on feature-based representations, separate one-class classifiers, and frequency-converted input. These methods struggle to detect anomalies in high-resolution time series that contain broader frequency information and complex time series shapes. Therefore, to address this issue, we propose the multisignal-based representation of complex time series for fault detection in industrial machinery. It includes a learning rule for the proposal that considers the purity and frequency diversity of the representation with a procedure to compute them efficiently, while the one-class decision function is part of the learned model. In detecting single- and multicomponent faults in reciprocating machinery, our method successfully solves the task, outperforming other classical and state-of-the-art approaches.

Echo state networks (ESNs) belong to the class of recurrent neural networks and have demonstrated robust performance in time series prediction tasks. In this study, we investigate the capability of different ESN architectures to capture spatial relationships in images without transforming them into temporal sequences. We begin with three pre-existing ESN-based architectures and enhance their design by incorporating multiple output layers, customising them for a classification task. Our investigation involves an examination of the behaviour of these modified networks, coupled with a comprehensive performance comparison against the baseline vanilla ESN architecture. Our experiments on the MNIST data set reveal that a network with multiple independent reservoirs working in parallel outperforms other ESN-based architectures for this task, achieving a classification accuracy of 98.43%. This improvement on the classical ESN architecture is accompanied by reduced training times. While the accuracy of ESN-based architectures lags behind that of convolutional neural network-based architectures, the significantly lower training times of ESNs with multiple reservoirs operating in parallel make them a compelling choice for learning spatial relationships in scenarios prioritising energy efficiency and rapid training. This multi-reservoir ESN architecture overcomes standard ESN limitations regarding memory requirements and training times for large networks, providing more accurate predictions than other ESN-based models. These findings contribute to a deeper understanding of the potential of ESNs as a tool for image classification.

Data scarcity in prognostic and health management research presents a significant challenge, often hindering the performance of supervised models due to the difficulty of acquiring diverse fault mode data during prolonged faultless operation. Conversely, nominal operating condition (NOC) data, including both healthy and varied faulty data, are more readily available due to predelivery inspection. Subsequently, we study this novel and unresolved NOC premise that leverages NOC data along with healthy data from other conditions to construct a fault diagnoser called Res-1D-bootstrap your own latent (BYOL) with the proposed probability distribution generalization strategy. The initial step involves a novel approach to the contrastive transformation optimization with the criteria based solely on similarity loss obtained in the training stage. We then pretrain the fault detector based on our NOC premise, followed by finetuning the network exclusively with NOC data. Given the novelty of our premise, there are few models for direct comparison. Thus, we contrast our approach with a supervised baseline, MoCo, an unoptimized equivalent algorithm, and an equivalent algorithm that solely employs NOC data for pretraining the feature extractor. Empirical results demonstrate our model's superior distribution generalization capabilities through the improved classification accuracy across different operating conditions.