This work investigates the model calibration, state estimation, and control of an Up-flow Anaerobic Sludge Blanket (UASB) reactor for the task of increasing its biogas production while complying with effluent quality requirements. We consider a dynamic model based on the Anaerobic Digestion Model No. 1 (ADM1), calibrated using experimental data from a blackwater treatment plant located in Helsingborg, Sweden. Using this model, we design a control solution composed of an operating point optimiser, defining the setpoints that achieve maximum biogas production, and a feedback controller, that tracks these setpoints while rejecting disturbances. For the feedback controller, we consider two alternative approaches: a multiloop Proportional-Integral-Derivate (PID) controller and an output-feedback Model Predictive Controller (MPC). The control solution is showcased on simulated experiments of operating the UASB reactor under different influent disturbance scenarios and compared with the nominal open-loop operation. Our results show that the controlled system is able to increase biogas production, while also improving effluent quality. Both control approaches have similar setpoint tracking performances, with the MPC being more economical and robust to influent disturbances.
Personal identification through biometric systems recognizes a person’s unique biological or behavioral characteristics. For high-security systems, user identification requires reliable and robust biometric methods. This work develops a method for footstep biometric recognition, based on pressure-sensor signals and convolutional neural networks, for access to restricted areas. The biometric acquisition device is a platform instrumented with fiber Bragg gratings to capture optical signals associated with the plantar pressure distribution. A preprocessing algorithm transforms the gratings’ wavelength shifts in the acquired signals into a plantar pressure map, separating the left and right feet into two smaller images. These small images are used to independently train two convolutional neural networks, one for each foot. After training, both CNNs are used as input to a fully connected neural network (FNN), forming a two-stage neural network (CNN-FNN) model whose weights are trained to identify a person correctly. This CNN-FNN biometric system identifies users from the complete plantar pressure map comprising both feet. The developed system achieves 98.67% accuracy, making it an efficient tool for inclusion in biometric security systems.
One of the main challenges in machine learning tasks is unlabeled data. Consequently, semi-supervised learning is an approach commonly used in labeling data tasks. The use of diffusion kernels is one of the powerful methods in this set of learning. This paper introduces Phylogram based Data Labeling (PhyDL), an innovative method that combines the benefits of phylogram analysis with Learning with Local and Global Consistency (LLGC). PhyDL leverages phylogram analysis to act as an aggregator of information from labeled samples, capable of correcting possible labeling errors. PhyDL constructs a phylogram using community detection algorithms, FastGreedy and Edge Betweenness, to identify and leverage the structure of the data, with FastGreedy emerging as the best-performing algorithm among those evaluated. We implemented PhyDL and evaluated its performance on the TwoMoons, MNIST, MUSK, HEART, and SONAR datasets. PhyDL outperformed LLGC on 4 out of 5 evaluated datasets, highlighting its effectiveness in semi-supervised learning scenarios. On the TwoMoons, MNIST, MUSK, and SONAR benchmarks, the method showed steady accuracy gains, with consistent improvements as the number of labeled samples increased. However, on the HEART dataset, standard LLGC achieved better results, indicating that PhyDL’s community-based diffusion can be limited by certain underlying spatial structures. Furthermore, the experiments revealed that the Edge Betweenness algorithm is unsuitable for these specific high-dimensional tasks. Ultimately, this study showcases PhyDL’s potential, when equipped with the FastGreedy algorithm, to significantly improve data labeling performance in semi-supervised scenarios, effectively leveraging the structural topology of complex datasets.
We propose using two deep autoencoder architectures, i.e. the Temporal Convolutional Network Autoencoder (TCN-AE) and Wasserstein Autoencoder with Generative Adversarial Network (WAE-GAN), to perform anomaly detection in the industrial context. In this work, we detect anomalies by introducing a feature engineering pipeline that structures univariate acoustic signals into sequential multivariate representations via Mel Frequency Cepstral Coefficients (MFCCs). The trained models were evaluated using 10-fold cross-validation, and their performances were assessed through an empirical analysis of hyperparameter sensitivity, mapping prominent structural configurations for each neural architecture. The optimized TCN-AE showed superior performance, achieving an average AUC-ROC of 0.8716 for pump devices and 0.7834 for fan devices, thereby surpassing state-of-the-art literature benchmarks, while also demonstrating robust performance for the slider device with an average AUC-ROC of 0.9646. Both architectures presented limitations with the valve device due to its highly transient behavior. The results achieved confirm the effectiveness of sequential reconstruction-based models for continuous machine operations while highlighting specific open challenges in acoustic industrial monitoring.
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.
