Loss of movement in the upper limb is a limitation that affects the lives of many individuals. This problem motivates the creation of several solutions, from intelligent biomechanical prostheses to new physical therapy approaches. However, one of the main limitations of these applications is in the detection and recognition of signal patterns for movement classification, considering both accuracy and computational complexity. The main hypothesis of this work is that extreme learning machines and echo state machines can combine acceptable accuracy with the low memory usage and processing typical of these machine learning models, in order to promote greater autonomy to users. In this work, we proposed a simple method for electromyography (EMG) classification, disregarding feature selection and post-processing algorithms. In this work, we propose the use of extreme learning machines and echo state machines as a low computational cost alternative for motion recognition from myoelectric signals. Other classifier models, such as the multilayer perceptron and linear discriminant analysis, were investigated. To validate the proposal, we used two datasets that represent two scenarios: an ideal environment, with no data on amputees; and a large dataset with myoelectric signals from patients with amputation. The results were evaluated as a function of accuracy and kappa index. Using an extreme learning machine, the results for the dataset of healthy individuals reached 99.10% accuracy, with a processing time per instance of 0.09 ms. Considering the amputee dataset, the best results were obtained with the use of extreme learning machine augmentation, with an accuracy of 75.60% and a kappa of 0.7507. We showed that Extreme Learning Machines without pre- and post-data processing is feasible for real-time application, presenting good performance and increased velocity when compared to usual algorithms. Also, data augmentation appears to be a good approach for robustly classifying unbalanced EMG datasets.

In December 2019, a member of the Coronaviridae family crossed the barriers between species and hit humans for the first time. Associated with Severe Acute Respiratory Syndrome (SARS) the virus was named SARS-CoV-2. The new coronavirus is responsible for 2019 coronavirus disease (Covid-19). Until August 20, the world has counted more than 22.5 million Covid-19 positive cases, about 14.4 million recovered cases, and more than 790 thousand deaths. In this study, we propose an automatic system for Covid-19 diagnosis using machine learning techniques and CT X-ray images named IKONOS-CT. The main idea is that healthcare professionals can upload the CT image into the system to be analyzed by an intelligent system. Then IKONOS-CT will provide a binary classification, differentiating Covid-19 patients from non-Covid-19 ones. For classification tasks we performed 25 experiments for the following classifiers: Multilayer Perceptron, Support Vector Machines, Random Tree and Random Forest, and Bayesian Networks. The best overall performance was found using Haralick as feature extractor and SVM with polynomial kernel of exponent 3. We found average results of accuracy of 96.994 ± 1.375%, kappa index of 0.940 ± 0.028, sensitivity/recall of 0.952 ± 0.024, precision of 0.987 ± 0.014, specificity of 0.987 ± 0.014, and area under ROC of 0.970 ± 0.014. By using a computationally low-cost method, based on Haralick extractor, it was possible to achieve high diagnosis performance. Indicating that an effective path for Covid-19 diagnosis may be found by combining AI for CT images classification and clinical analysis.

Breast cancer is one of the most prevalent types of cancer and the deadliest form of cancer among women. The detection and early diagnosis of cancer are of fundamental importance to increase the possibility of treatment effectiveness, reducing mortality rates. Breast thermography produces high-resolution infrared images that show metabolic changes resulting from the appearance of altered cells in breast tissue. Despite being a promising technique, the interpretation of thermography images is often difficult. Pattern recognition techniques have the potential to work around this problem, helping to extract more useful information from these images. In this work, we propose the selection of attributes based on the dialectic method of optimization in breast thermography, aiming to simplify the classifiers and increase the potential of generalization to support the diagnosis of breast lesions. Through the proposed attribute selection technique, it was possible to simplify the classifier architectures, reducing the dimensionality of the attribute vectors by about 50% with a low impact on the classification’s correct rates, with a reduction of around 3.72%. The proposed method is a promising technique for reducing attributes, with significant accuracy values being obtained using only 84 of the 168 attributes originally extracted. This shows the importance of this step for the use of breast thermography as an auxiliary technique for the diagnosis of breast cancer.

The life expectancy of the population in the most developed countries is growing every day and, consequently, there is an increase in various age-related diseases. In Brazil, just over 1.1 million people have Alzheimer’s disease (AD). In 2019, according to the World Health Organization, Alzheimer’s disease and other dementias were the third leading cause of mortality in the Americas and Europe. Despite being a degenerative and irreversible disease, if diagnosed early, treatments can be performed to slow the progression of symptoms and ensure a better quality of life for the patient. Most papers that study Computational Intelligence solutions to support diagnosis follow an approach based on neuroimaging evidence. In addition to this, another approach that has been gaining prominence is biochemical and molecular analysis. Following this approach, Ray et al., Ravetti & Moscato and Dantas & Valença carried out studies with classifiers from statistics or Computational Intelligence to support the early diagnosis of the disease. The work was carried out from a dataset with values of 120 blood proteins. Through this, they were able to classify whether or not the patient could be diagnosed with AD. As a result, Ray et al., Ravetti & Moscato and Dantas & Valença obtained an average accuracy rate of 89%, 93% and 94.34%, respectively. Thus, this work aims to use a traditional approach with a proposed Multilayer Perceptron Artificial Neural Network model to perform the early diagnosis of a patient with or without AD and compare the results obtained with the results of the related works cited. In addition, this work has as main objective to evaluate the potential of using synthetic data generated using a Generative Adversarial Network in the training and tests of the proposed classification model.

In late 2019, the SARS-Cov-2 spread worldwide. The virus has high rates of proliferation and causes severe respiratory symptoms, such as pneumonia. There is still no specific treatment and diagnosis for the disease. The standard diagnostic method for pneumonia is chest X-ray image. There are many advantages to using Covid-19 diagnostic X-rays: low cost, fast and widely available. We propose an intelligent system to support diagnosis by X-ray images.We tested Haralick and Zernike moments for feature extraction. Experiments with classic classifiers were done. Support vector machines stood out, reaching an average accuracy of 89.78%, average recall and sensitivity of 0.8979, and average precision and specificity of 0.8985 and 0.9963 respectively. The system is able to differentiate Covid-19 from viral and bacterial pneumonia, with low computational cost.

One of the main challenges in the handwriting recognition area lies in identifying complete lines of handwritten text. In this paper, we propose a handwriting recognition system based on a deep multidimensional long-short-term memory (MDLSTM) network within a hybrid hidden Markov model framework. The MDLSTM architecture was elaborated to enhance the recognition performance and decrease the recognition time. Accordingly, we present modifications regarding the layers order and the number of pooling layers compared to a standard MDLSTM model. Since the results reported in the literature for deeper MDLSTM architectures relies on optimizing the network width with a fixed depth, we investigate the trade-off between both these properties to obtain an optimal topology. The system was evaluated with English handwritten text lines from the IAM database and the experiments demonstrated that the proposed MDLSTM architecture was able to maintain a robust recognition performance (around 3.6% CER and 10.5% WER) and present significant speedups, approximately 48% and 32% faster than the state-of-the-art MDLSTM optical model, regarding the learning and classification times, respectively. The full system including a decoder with linguistic knowledge presents competitive results with the state-of-the-art.

Since 2013, the São Francisco River has being through a low hydraulicity period. In other words, the rain intensity is below average. Consequently, it has being necessary to operate at a minimal flow rate. It is far below the ones established at the operation licence, which is 1300 m $$^3$$ /s. Due to this hydraulic crisis, the actual operational flow rate is 700 m $$^3$$ /s at São Francisco River, characterizing this situation as critical. In this work, it was proposed to use Reservoir Computing (RC), Long Short Term Memory (LSTM) and Deep Learning to predict Sobradinho's flow rate for 1, 2 and 3 months ahead using macroclimatic variables. After having the results for each one of them, a comparison was made and statistical tests where executed for evaluation.

Some databases used in computer vision problems have a few number of data points. It makes harder to the classifier to increase its generalization capability. One strategy applied to solve this issue is data augmentation. This solution aims to generate more data to improve the performance of the classifier. In this paper, we propose a model that uses images produced by an Extreme Learning Machine Autoencoder (ELM-AE) to make data augmentation. We selected the autoencoder approach since it is more straightforward and efficient than other data augmentation strategies. We evaluate our proposal in a facial expression recognition problem using the Japanese Female Facial Expression (JAFFE) database, and we assess the impact of the data augmentation in the performance considering K-Nearest Neighbors (KNN), Support Vector Machines (SVM) and Convolutional Neural Network (CNN). The obtained results show that our approach is an appropriate alternative for data augmentation tasks, also reaching better results when compared to other common strategies in most of the cases.

The importance of organizing medical images according to their nature, application and relevance is increasing. Furhermore, a previous selection of medical images can be useful to accelerate the task of analysis by pathologists. Herein this work we propose an image classifier to integrate a CBIR (Content-Based Image Retrieval) selection system. This classifier is based on pattern spectra and neural networks. Feature selection is performed using pattern spectra and principal component analysis, whilst image classification is based on multilayer perceptrons and a composition of self-organizing maps and learning vector quantization. These methods were applied for content selection of immunohistochemical images of placenta and newdeads lungs. Results demonstrated that this approach can reach reasonable classification performance.

Smart Houses and Internet of Things (IoT) are two present tendencies in our days. Due to these technologies, the existent types of equipment in a smart house (sensors, thermostats, and video cams) allow us to analyze and collect data from a person's daily activities and use it in the field of anomaly detection. Therefore, noninvasive monitoring techniques can be applied to people's residences. When focusing on the elderly population, this type of approach can be used to detect and report a fall, decreasing the costs of monitoring these individuals. This paper uses images from a Microsoft Kinect cam, accelerometers' data, digital image processing and computer vision techniques to make a comparative study between different supervised classifiers and statistic approaches when they are being used in the fall detection problem. The results show that some of the tested classifiers are efficient in this task, reaching an accuracy of 96.67% and 98.79%.

Streamflow forecasting is a fundamental tool in water resource studies. If information on the nature of the inflow is determinable in advance, then a given reservoir can be operated by some decision rule to minimize downstream flood damage and maximize the generated power with low costs. However, traditional methods such as linear time series models do not model the series properly, ignoring its dynamical behavior. This paper provides a method based on the Reservoir Computing (RC) technique combined with trend information extracted from the series for short-term streamflow forecasting. The model was tested in five hydroelectric plants located in different river basins in Brazil. Experimental results show that the proposed method is able to achieve better generalization performance than the traditional methods.

Um dos principais desafios da atualidade é a crescente demanda de energia mundial e, a fim de suprir essa necessidade, as fontes de energia mais utilizadas são o petróleo, gás natural e carvão mineral. O grande problema com essas fontes deve-se ao fato de, além de serem extremamente poluidoras, elas são de origem não renovável. Por causa disso, energias renováveis estão se tornando cada vez mais essenciais para a humanidade e entre elas, está o vento e sua escolha é uma das mais promissoras. Os parques eólicos tem seu potencial diretamente ligado à potência do vento, sendo necessária boas estimativas dessa variável para a construção de estratégias e planejamentos eficientes. Entretanto, essa tarefa apresenta grandes dificuldades devido às complexas caracterı́sticas do vento, como a alta variabilidade de sua velocidade e direção. Este trabalho tem como objetivo utilizar a técnica de Reservoir Computing para a previsão da potência do vento e, pelo fato da Multi-Layer Perceptron ser a mais utilizada para tal fim, realizar a comparação entre esses dois tipos de Redes Neurais Artificias e analisar qual possui o melhor desempenho.

The increasing use of wind power as source of electricity motivates a continuous improvement of the accuracy of wind power forecasts. There is a considerable value in optimizing forecasts systems to provide the best performance in an environment where the wind power production increases and/or decreases by a large amount over a short period of time. This paper presents a method that uses Reservoir Computing to forecast variations in energy production in wind farms, or ramp events. This method is compared with two other approaches: one that uses a MLP network and the other is based in persistence. The tests were performed and the results are given for a real case. Keywords—wind power forecast; reservoir computing; ramp events; phase errors.

Artificial Neural Networks have been widely adopted to tackle problems of classification and prediction. Many applications of this technique presented more accurate results when compared to traditional statistical techniques. In order to solve general problems, one needs to use the Multilayer Perceptron. The most common algorithm to train this type of neural network is called Backpropagation. However, this algorithm presents some drawbacks, such as slow convergence and the possibility to get stuck in local minima. This paper proposes to apply some recently proposed adaptive variations of particle swarm based algorithms to train neural networks. The ClanAPSO algorithm presented better results in the task of training the neural network, although its duration and number of iterations until convergence increased significantly.

Neste trabalho é apresentado um método para classificação automática de distúrbios elétricos baseado em redes neurais artificiais, utilizando no treinamento sinais reais de tensão coletados através de registradores digitais de perturbação existentes no sistema de monitoração da CHESF. O sinal de tensão perturbado é coletado a uma taxa de amostragem de 128 amostras/ciclo na freqüência de 60Hz durante 14 ciclos. O sinal de tensão real coletado é processado em duas etapas: i) inicialmente é decomposto através da transformada wavelet até o quinto nível de resolução, extraindo suas principais características; ii) em seguida os coeficientes wavelets são processados via Análise de Componentes Principais que opera projetando os dados linearmente em um subespaço de menor dimensão. A classificação realizada pela combinação de três redes MLP com diferentes arquiteturas é baseada nas características extraídas a partir do pré-processamento do sinal de tensão. O algoritmo Resilient Backpropagation foi utilizado no treinamento das redes. Na combinação das redes, em cada um dos seis nós de saída, aplicou-se a média entre as três saídas das redes individuais. A decisão final do classificador corresponde à saída combinada de maior valor. Os resultados são bastante promissores para os cinco tipos de distúrbios testados – Afundamentos e Elevação de tensão, Harmônicos, Transitórios oscilatórios, Interrupção; e para ausência de distúrbio. 1. Introdução