Eréndira Rendón Lara

Surface electromyography (sEMG) signals acquired with low-cost sensors tend to exhibit variable degradation that can compromise the reliability of myoelectric control systems outside controlled conditions. This work presents an adaptive processing pipeline for sEMG signals composed of three chained stages: a motor-intent classifier based on dilated temporal convolutional networks, whose function is to determine whether a signal window contains muscle activity associated with a voluntary gesture; a dual-output quality assessor that estimates a continuous score and a binary acceptability label, aimed at deciding whether the signal can be used directly or requires intervention; and a convolutional autoencoder that recovers the morphology of degraded windows before they are used in prosthetic control. The decision policy for reconstruction operates on two independent thresholds and classifies each window into one of three states: signal discard, direct acceptance, or active restoration. The models within the proposed architecture are trained on the public NinaPro DB1, DB3, and DB10 datasets and validated without recalibration on signals recorded from two participants with transradial amputation over three to four weekly sessions. The results show that the pipeline correctly handles signal profiles with opposing characteristics. Inference latency remained below 5 ms at the 50th percentile across all scenarios, consistent with real-time operation. All inference was executed on a host computer; a prototype using an Arduino UNO R4 WiFi as a peripheral interface was built to display the pipeline’s decisions on physical hardware and to confirm that the serial-communication link does not introduce additional latency, not to perform on-board inference. A downstream evaluation further showed that, while restoration improved signal-level fidelity metrics, it did not translate into improved motor-intent classification accuracy relative to using the degraded signal directly, a limitation discussed explicitly in this work.

Surface electromyography (sEMG) signals acquired with low-cost sensors tend to exhibit variable degradation that can compromise the reliability of myoelectric control systems outside controlled conditions. This work presents an adaptive processing pipeline for sEMG signals composed of three chained stages: a motor-intent classifier based on dilated temporal convolutional networks, whose function is to determine whether a signal window contains muscle activity associated with a voluntary gesture; a dual-output quality assessor that estimates a continuous score and a binary acceptability label, aimed at deciding whether the signal can be used directly or requires intervention; and a convolutional autoencoder that recovers the morphology of degraded windows before they are used in prosthetic control. The decision policy for reconstruction operates on two independent thresholds and classifies each window into one of three states: signal discard, direct acceptance, or active restoration. The models within the proposed architecture are trained on the public NinaPro DB1, DB3, and DB10 datasets and validated without recalibration on signals recorded from two participants with transtibial amputation over four weekly sessions. The results show that the pipeline correctly handles signal profiles with opposing characteristics. Inference latency remained below 5 ms at the 50th percentile across all scenarios, consistent with real-time operation. An integrated prototype using an Arduino UNO R4 WiFi was used to evaluate the results, and validate the feasibility of the system on physical hardware.

In Latin American and Caribbean States the verbal violence against women on social networks, such as Twitter, is a serious threat that has been addressed through the implementation of social norms, public policies, and social movements. Nevertheless, a challenge is the effective and automatic real-time detection of violent tweets. In this sense, traditional machine learning algorithms have been proposed to tackle social issues where the training process is performed in a static manner. However, considering that Twitter is a dynamic environment where a vast of tweets are generated each second, it requires powerful machine learning algorithms that could exploit this pool of unlabeled data to be incorporated into the model through continuous updates. This paper explores an active learning method based on uncertainty sampling, which identifies the most confusing tweets to be labeled by an expert in real-time. This focused selection prioritizes which data can be used to train a multilayer perceptron that can achieve a better performance with fewer training samples. Experimental results show that including new samples yields promising results, increasing the AUC from 0.8712 to 0.8833.

Dyes are organic compounds capable of transmitting their color to materials, which is why they are widely used, for example, in textile fibers, leather, paper, plastic, and the food industry. In the dying process, measuring the dye’s content is extremely important to evaluate the process efficiency and minimize the dye’s discharge in wastewater, but most of the time, dyes are present in multi-component mixtures; hence, quantification by spectrophotometric methods presents a great challenge because the signal obtained in the measurement overlaps the components in the mixture. In order to overcome this issue, the use of the high-performance liquid chromatography (HPLC) method is recommended; however, it has the disadvantage of being an expensive technique, complex, and requiring excessive sample preparation. In recent years, some direct spectrophotometric methods based on multivariate regression algorithms for the quantification of dyes in bicomponent mixtures have been reported. This study presents a new framework that uses a combined ANN and principal component analysis (PCA) model for the determination of the concentration of three dyes in aqueous mixtures: Tartrazine (TZ), Amaranth Red (AR), and Blue 1 CFC (B1) dyes. The PCA–ANN model was trained and validated with ternary mixture samples of TZ, AR, and B1, and with known different compositions, spectra absorbance samples were measured in a UV-Vis spectrophotometer at wavelengths between 350–700 nm with intervals of 1 nm. The PCA–ANN model showed a mean absolute prediction error and correlation coefficient (r²) of less than 1% and greater than 0.99, respectively. The results demonstrate that the PCA–ANN model is a quick and highly accurate alternative in the simultaneous determination of dyes in ternary aqueous mixtures.

In machine learning and data mining applications, an imbalanced distribution of classes in the training dataset can drastically affect the performance of learning models. The class imbalance problem is frequently observed during classification tasks in real-world scenarios when the available instances of one class are much fewer than the amount of data available in other classes. Machine learning algorithms that do not consider the class imbalance could introduce a strong bias towards the majority class, while the minority class is usually despised. Thus, sampling techniques have been extensively used in various studies to overcome class imbalances, mainly based on random undersampling and oversampling methods. However, there is still no final solution, especially in the domain of multi-class problems. A strategy that combines density-based clustering algorithms with random undersampling and oversampling techniques is studied in this work. To analyze the performance of the studied method, an experimental validation was achieved on a collection of hyperspectral remote sensing images, and a deep learning neural network was utilized as the classifier. This data bank contains six datasets with different imbalance ratios, from slight to severe. The experimental results outperform the classification measured by the geometric mean of the precision compared with other state-of-the-art methods, mainly for highly imbalanced datasets.

This paper presents a performance comparison between three models of pre-trained CNN networks (VGG16, ResNet50, and MobileNet) in detecting physical violence against women in video. To carry out the classification of images that include physical violence against women and those that do not, 2 800 images (1 400 violent and 1 400 non-violent) were collected from a public dataset and subsequently divided into training (1 200 images), validation (1 000 images) and test (600 images). To evaluate their performance, accuracy values for each model were considered, positioning the MobileNet network as the best-performing classifier for this classification task with 89% accuracy.