The growing integration of sensors and embedded devices into distributed Internet of Things (IoT) systems has increased the demand for real-time data collection and processing solutions. However, the high volume and frequency of sensor data create challenges related to storage, transmission, and response latency, especially in resource-constrained environments. In this context, locally executed compression techniques, aligned with the Tiny Machine Learning (TinyML) paradigm, become differentiators for enabling embedded applications. Thus, this work proposes an autoencoder-based approach for efficiently compressing sensor data on edge devices. Three autoencoder variants (feed-forward, sparse, and contractive) are evaluated, combined with symmetric and asymmetric architectures, considering criteria such as compression ratio, information preservation, and embedded execution feasibility. For practical validation, a case study was conducted using vehicular data collected via the OBD-II interface, where the selected models were deployed on the OBDII Edge Freematics One+ device. The results show that the models could reduce data dimensionality with minimal information loss, maintain competitive performance on discriminative tasks, and exhibit inference times compatible with real-time applications. Autoencoders represent a viable neural compression solution for IoT environments, potentially applicable to various embedded sensing scenarios.

Kolmogorov–Arnold Networks (KANs) represent a promising machine learning architecture that leverages univariate functional decomposition to model complex phenomena using compact and interpretable structures. These characteristics make KANs especially attractive for deployment in TinyML environments, where memory, processing power, and energy consumption are strictly constrained. This paper evaluates the feasibility and trade-offs of using a KAN model to estimate the State of Charge (SoC) in electric vehicle batteries. We design a KAN tailored for embedded systems and compare its performance with a conventional Multilayer Perceptron (MLP) baseline under identical training and deployment conditions. Our evaluation includes predictive accuracy, training cost, model size, inference speed, and energy consumption on microcontrollers. Results show that the KAN model achieves a nearly ten times smaller memory footprint than the MLP (693 bytes vs. 6807 bytes) and maintains comparable energy consumption and inference speed across different embedded platforms. Although the MLP outperforms the KAN during training with faster convergence and lower energy requirements, the KAN demonstrates competitive predictive performance at inference time while significantly reducing deployment costs in terms of memory usage and energy efficiency at the edge. Furthermore, the KAN model produces symbolic mathematical expressions, offering direct interpretability and facilitating analytical validation — a critical advantage for embedded battery diagnostics and safety-critical applications.

The advancement of Industry 4.0 has intensified the demand for intelligent systems that can efficiently access and interpret technical information critical to industrial operations. However, recovering knowledge from extensive and complex technical documentation remains a significant challenge. This study examines the effectiveness of various Retrieval-Augmented Generation (RAG) strategies, combined with different Large Language Models (LLMs), in extracting and generating answers from industrial technical documents. A case study was conducted based on PROFIBUS, a widely adopted digital communication protocol in automation networks, with technical documents organized into categories for engineers and developers. Twenty questions of varying complexity were formulated, and responses were generated using three RAG strategies (Basic, Decomposition, and HyDE) combined with two LLMs (Gemma 3 and GPT-4o-mini). The outputs were compared against reference answers generated by the NotebookLM system and evaluated using automatic metrics, including ROUGE, METEOR, BERTScore, and MATTR. The results indicate that the Decomposition and HyDE strategies achieved superior semantic similarity scores when combined with more capable models. That model’s performance varied depending on the complexity of the document profile. These findings highlight the importance of tailored RAG strategies in enhancing intelligent information retrieval in industrial domains, which supports safer and more efficient operational environments.

The Internet of Things (IoT) generates massive multivariate time series data requiring real-time anomaly detection and correction for reliable monitoring. This challenges resource-constrained embedded devices due to conventional offline training and batch processing. To address this, we propose two algorithms derived from the TEDARLS framework: Multivariate Sequential TEDA (MST) and Multivariate Parallel TEDA (MPT). Derived from the TEDARLS framework, both enable on-device detection and correction of multivariate anomalies within TinyML constraints. A case study with real vehicular sensor data demonstrated low inference times and consistent embedded behavior. Supervised metrics were only assessed on synthetic data. MPT, though more sensitive, introduced greater signal distortions and required significantly longer processing times. Overall, MST demonstrated superior stability and suitability for real-time anomaly correction in resource-constrained IoT environments. This approach addresses an important gap in embedded analytics for IoT by enabling lightweight, accurate, and autonomous anomaly detection and correction at the edge.