In pervasive healthcare monitoring, activity recognition is critical information for adequate management of the patient. Despite the great number of studies on this topic, a contextually relevant parameter that has received less attention is intensity recognition. In the present study, we investigated the potential advantage of coupling activity and intensity, namely, Activity-Intensity, in accelerometer data to improve the description of daily activities of individuals. We further tested two alternatives for supervised classification. In the first alternative, the activity and intensity are inferred together by applying a single classifier algorithm. In the other alternative, the activity and intensity are classified separately. In both cases, the algorithms used for classification are k-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Random Forest (RF). The results showed the viability of the classification with good accuracy for Activity-Intensity recognition. The best approach was KNN implemented in the single classifier alternative, which resulted in 79% of accuracy. Using two classifiers, the result was 97% accuracy for activity recognition (Random Forest), and 80% for intensity recognition (KNN), which resulted in 78% for activity-intensity coupled. These findings have potential applications to improve the contextualized evaluation of movement by health professionals in the form of a decision system with expert rules.

Summary: This research addresses one of the most critical challenges in oilwell drilling: the early detection of kick anomalies. A kick anomaly is characterized by an undesirable influx of formation fluids into the wellbore, which can lead to severe safety hazards, making its timely identification essential to ensure the safety of all personnel involved and to maintain operational continuity. As kick cases are rare and operational methods often overestimate their occurrence, deep learning anomaly detection techniques are valuable for enhancing efficiency and minimizing false alarms. By using different autoencoder (AE) models, especially with convolutional networks, the aim of this study is not only to improve the accuracy of kick detection but also to anticipate the detection significantly. Based on the analyses of kick anomalies in deepwater offshore wells located in pre-salt Brazilian oil fields, we used real-time drilling data from correlation wells without anomalies to train different AE architectures. The data set comprised 730 hours of historical real-time monitoring records from eight offshore drilling operations across two different oil fields, during which two real kicks were documented. The results indicate that the 1D convolutional neural network AE (1D CNN-AE) outperforms five other well-established AE models reported in the literature for detecting kicks during offshore drilling operations. The 1D CNN-AE model demonstrated high efficacy in the early identification of kick events, achieving a recall rate of 100%. Furthermore, it yielded a 31% reduction in the historical false kick rate (FKR), indicating a substantial improvement in event classification accuracy. In addition to its predictive reliability, the model provided a mean early detection time (EDT) of approximately 100 minutes prior to the corresponding operational reports, reinforcing its potential as a robust tool for real-time decision support in well control operations. This study presents three original contributions: (i) a framework outlining the neural network techniques used in the literature to address the kick detection problem; (ii) a comprehensive investigation and preprocessing of kick event data recorded between 2018 and the first semester of 2025 in specific deepwater offshore wells in Brazil; and (iii) a prototype based on a 1D CNN-AE architecture.

Goal: Industry 4.0 enables the design of new models for process monitoring in which sensors, analyzers, and controls are positioned at different points in the process. The goal of this work is to present the modeling and control of a three-phase production separator with hydrocyclones to treat produced water on an oil platform of mature fields. Design / Methodology / Approach: The Methodology or approach used was to develop a model for the primary separator that allows its operation by means of a controller (fuzzy and PI) to manipulate the flow of discarded water, acting indirectly in the oil-in-water measure. Results: The results showed the consistency of the model for open loop simulations and the effectiveness of the controllers to comply with the discarding requirements for the closed-loop simulations. Limitations of the investigation: The limitation of the model and developed controllers is that they are only applied to platforms where water production exceeds the separators discharge capacity and the exceeding water can be offloaded to another equipment or to another platform. Practical implications: The main practical implications of this study are to maximize the flow of discarded water on mature field platforms, which produces elevated amounts of water, to conform the total oil and grease to the local law regulations. Additionally, it also increases oil production, with a higher limit of water production. Originality / Value: Compared to the previous authors, where the models of discarded flow is a function of the water-oil interface, this work developed a model that allowed the flow of the discarded water to be a function of its quality.

Human Activity Recognition (HAR) plays a crucial role in applications such as health monitoring, physical training, and therapeutic support. Smartphones equipped with embedded sensors provide a convenient platform for these applications. However, running deep learning models continuously can drain device energy autonomy. Despite recent advances, few studies have explored how executing deep learning-based HAR models across different computational environments affects energy consumption, especially when combined with strategies that leverage application-level contextual information. This work develops and evaluates optimized deep learning models, including variants with early exits, in both local and remote execution scenarios. It also incorporates an adaptive sampling strategy that dynamically adjusts sensor reading frequency according to movement intensity (contextual information). Results show that the proposed approach reduces energy consumption and inference time while preserving model accuracy.