Carlos Alberto Astudillo

Reliable discrimination of Volcano-Tectonic (VT) and Long-Period (LP) seismic signals is central to volcanic surveillance, yet operational deployment is hindered by pronounced inter-volcano domain shift, scarce annotated catalogues, and heterogeneous acquisition instrumentation. We present a self-supervised trimodal deep learning framework that jointly encodes spectrograms, raw one-dimensional waveforms, and physical tabular features, and we evaluate its capacity for cross-domain generaliza-tion across three leave-one-volcano-out transfer settings among Galeras, Cumbal, and Chiles-Cerro Negro (Colombia). Beyond classification accuracy, we systematically characterise the geometry of the learned latent space using class- and domain-separability landscapes projected into two dimensions, quantifying silhouette coefficients, Davies–Bouldin indices, separation indices, and decision-boundary alignment for every ablation variant. The trimodal framework (SSL Trimodal) attains the highest bootstrap balanced accuracy on the Galeras+Cumbal→Chiles-Cerro Negro transfer (0.9080±0.0072) and the most favourable class-separation index (Sep-Idx_2D = 2.06) among all eight evaluated variants in that scenario. It further achieves the best class-separation index (1.72) even on the most divergent transfer setting, Galeras+Chiles-Cerro Negro→Cumbal, despite not attaining the highest accuracy there. This dissociation between point accuracy and embedding quality is, we argue, the more in-formative signal for assessing cross-domain robustness than accuracy alone. We report this result alongside its limitations transparently: the CORAL covariance-alignment baseline outperforms the trimodal framework on two transfer settings with comparable mean accuracies ( 0.81), underscoring that discriminative contrastive pretraining and second-order domain alignment address complemen-tary, rather than redundant, failure modes. To our knowledge, this is the first trimodal self-supervised framework reported for cross-domain generalization in seismo-volcanic signal classification.

Environmental monitoring systems rely on large-scale Internet of Things (IoT) deployments to capture continuous data streams describing air quality, water resources, and ecological dynamics. The reliability of these systems depends not only on data availability but also on the integrity, transparency, and long-term trustworthiness of the collected information. However, conventional data management approaches lack mechanisms to guarantee tamper-resistant records while maintaining operational sustainability. This work presents a sustainable approach to IoT data integrity by integrating distributed ledger technologies (DLT) into environmental monitoring infrastructures. Rather than focusing solely on transaction-level performance, the proposed approach considers system-level sustainability, including resource constraints at the edge, continuous data generation, and initial deployment feasibility under realistic operational conditions. An end-to-end architecture combining IoT sensing devices, edge processing, and IOTA Rebased as the notarization backend is designed to support continuous data anchoring without disrupting real-time monitoring requirements. Results from a 24-hour empirical study (2,342 notarization events) show a mean end-to-end latency of 2,618 ms and a success rate of 98.94%. The study highlights how careful system design enables a balance between trust, scalability, security, and sustainability, providing a practical pathway for deploying trustworthy environmental monitoring systems at scale.

Distributed ledger technologies (DLT) can enhance trust and auditability in the Internet of Things (IoT). Among them, IOTA has been specifically designed to support machine-to-machine interactions and IoT data anchoring through scalable DLT architectures. However, their integration with Low-Power Wide-Area Networks (LPWANs) remains limited due to device constraints, strict timing requirements, and the operational costs of on-chain transactions. The transition from the fee-less Stardust to the fee-based IOTA Rebased model introduces explicit transaction costs, questioning the viability of continuous IoT data anchoring. IOTA provides a suitable platform to examine the challenges of integrating distributed ledger technologies with LPWAN-based IoT systems. Its transition to a fee-based execution model raises important questions regarding cost predictability and performance in continuous data anchoring scenarios, particularly under the constraints of resource-limited and latency-sensitive environments. This article investigates the practicality of the execution and payment model introduced by IOTA Rebased for IoT scenarios requiring continuous data notarization. We provide an empirical evaluation of continuous IoT data notarization on the public IOTA Rebased Mainnet and characterize the performance implications on edge-oriented deployments, including resource-constrained and resource-rich devices. We implement a notarization oracle that ingests LoRaWAN uplinks from The Things Network (TTN), canonicalizes payloads, generates SHA-256 commitments, and records them on-chain through reusable notarization objects. The oracle enables continuous anchoring of IoT telemetry while minimizing transaction overhead through object reuse. Two 24-h experimental campaigns compare a notarization oracle on resource-constrained and resource-rich hardware under periodic workloads. Results show consistent steady-state gas consumption for UPDATE operations, indicating that object reuse enables stable on-chain cost behavior in IOTA Rebased regardless of the deployment platform. From a performance perspective, both environments achieve stable execution; however, the resource-constrained edge deployment exhibits higher median and tail latency, alongside tighter memory margins compared to the resource-rich centralized baseline. These findings confirm the feasibility of deploying notarization services on constrained edge infrastructure under the new fee-based model.

Cross-domain transfer of seismo-volcanic classifiers across geologically distinct volcanic systems remains a critical challenge: models trained on one volcano degrade when deployed at a new site without target-domain labels. This work proposes the Phase-Modulated Hypergraph Fusion (PMHF) model, a trimodal deep architecture that jointly exploits CLIP-encoded spectrograms, waveform images, and physics-informed seismic features to discriminate Long-Period (LP) from Volcano-Tectonic (VT) events in an adaptation domain setting. PMHF integrates three novel modules: a Spectral Calibrated Physics Gate (SCPG) that recalibrates tabular features via class-conditional resonance matrices; Dynamic Hypergraph Convolutions (DHGC) that propagate visual features through probability-conditioned hyperedges; and a Trimodal Episodic Cross-Attention (TECA) mechanism that fuses all pairwise modal interactions through persistent memory. Three leave-one-volcano-out scenarios evaluated across Galeras, Cumbal, and Chiles-Cerro Negro (bootstrap N=1000) show that PMHF achieves mean balanced accuracy 0.765, peaking at 0.868±0.009 (MCC =0.743±0.017) for Galeras+Cumbal→Chiles-Cerro Negro, where bimodal baselines plateau below 0.659 and tabular models exhibit systematic class inversion (MCC =−0.405). Domain adaptation diagnostics confirm that trimodal fusion achieves the tightest and most stable source-target feature alignment, enabling robust volcanic monitoring without retraining.

Massive Machine-Type Communications (mMTC) scenarios in Cellular Internet of Things (CIoT) networks may generate peaks of simultaneous access attempts, making the random access (RA) procedure a critical bottleneck. In such conditions, preamble collisions increase contention, retransmissions, and access delay. Recent supervised approaches based on power-delay-profile (PDP) features have shown promising performance for early preamble collision detection, but they depend on labeled datasets. In this domain, labeled data is costly and may not be available (e.g., in new network deployments) because collision labels require detailed simulation and protocol-level verification. This paper investigates self-supervised learning (SSL) to reduce label dependence while improving downstream collision classification and cross-scenario robustness. A Transformer-based encoder is pre-trained on unlabeled PDP sequences, and representations learned are then transferred via fine-tuning with a small labeled subset, employing a simple multilayer perceptron (MLP) classifier as a prediction head. Two masking-based SSL pretext tasks are evaluated, namely random token masking and contiguous masked-span prediction, using the datasets with varying characteristics, such as channel type, speed, and coverage radius. Simulation results show that SSL improves the downstream task performance in all same-scenario evaluations and in five of the six cross-scenario transfer pairs, with the largest gains under domain mismatch. These findings indicate that SSL reduces dependence on labeled data while improving transferability across different radio channel conditions.

O crescimento de aplicações de Internet das Coisas (IoT) em cenários de comunicações massivas do tipo machine-type communications (mMTC) impõe desafios significativos aos procedimentos de acesso aleatório em redes celulares. Em particular, a elevada ocorrência de colisões no canal de acesso aleatório (PRACH) compromete a eficiência do sistema, resultando em aumento da latência de acesso e no número de retransmissões. Este trabalho propõe uma abordagem integrada para detecção explícita de colisões baseada em aprendizado de máquina, combinada com políticas de priorização conscientes de colisão no procedimento de acesso aleatório. O modelo proposto é treinado a partir de dados simulados e otimizado para execução em ambientes de borda por meio de técnicas de quantização. Avaliações experimentais em cenários mMTC com até 30 mil dispositivos simultâneos demonstram reduções significativas na latência de acesso e no número médio de retransmissões, evidenciando o potencial da abordagem para redes celulares IoT de alta densidade.

Preamble collision in the random access channel (RACH) is a major bottleneck in massive machine-type communication (mMTC) scenarios, typical of cellular IoT (CIoT) deployments. This work proposes a machine learning-based mechanism for early collision detection during the random access (RA) procedure. A labeled dataset was generated using the RA procedure messages exchanged between the users and the base station under realistic channel conditions, simulated in MATLAB. We evaluate nine classic classifiers -- including tree ensembles, support vector machines, and neural networks -- across four communication scenarios, varying both channel characteristics (e.g., Doppler spread, multipath) and the cell coverage radius, to emulate realistic propagation, mobility, and spatial conditions. The neural network outperformed all other models, achieving over 98\% balanced accuracy in the in-distribution evaluation (train and test drawn from the same dataset) and sustaining 95\% under out-of-distribution evaluation (train/test from different datasets). To enable deployment on typical base station hardware, we apply post-training quantization. Full integer quantization reduced inference time from 2500 ms to as low as 0.3 ms with negligible accuracy loss. The proposed solution combines high detection accuracy with low-latency inference, making it suitable for scalable, real-time CIoT applications found in real networks.

In 5G networks, base station (BS) disaggregation and new services present challenges in radio access network (RAN) configuration, particularly in meeting their bandwidth and latency constraints. The BS disaggregation is enabled by functional splitting (FS), which distributes the RAN functions in processing nodes and alleviates latency and bandwidth requirements in the fronthaul (FH). Besides network performance, energy consumption is a critical concern for mobile network operators (MNO), since RAN operation constitutes a major portion of their operational expenses (OPEX). RAN configuration optimization is essential to balance service performance with cost-effective energy consumption. In this paper, we propose a mixed-integer linear programming (MILP) model formulated with three objective functions: (i) minimizing fronthaul (FH) latency, (ii) minimizing energy consumption, and (iii) a bi-objective optimization that jointly balances both latency and energy consumption. The model determines the optimal FS option, RAN function placement, and routing for eMBB, URLLC, and mMTC slices. Although prior studies have addressed RAN configuration either from an energy minimization or latency reduction perspective, few have considered both aspects in realistic scenarios. Our evaluation spans different topologies, accounts for variations in aggregated gNB demand, explores diverse FS combinations, and incorporates Time Sensitive Networking (TSN) modeling for latency analysis, as it is also crucial in RAN performance. Given that MILP's execution time can be significant, we propose a heuristic algorithm that adheres to RAN constraints. Our results reveal a trade-off between latency and energy consumption, highlighting the need for dynamic RAN reconfiguration. These insights provide a foundation to optimize existing and future RAN deployments.