Search

Recently, the memory mechanism has been widely implemented in target tracking. However, these trackers hardly balance the stability of long-term memory with the plasticity of short-term memory through an elegant and efficient mechanism. A residual memory inference network (RMIT) is proposed to exploit the history of target states and last visual features. Specifically, RMIT consists of a base layer and a residual memory layer by synergizing short-and long-term memories. The base layer can be regarded as Discriminative Correlation Filter (DCF) reformulation that maintains the short-term memory to accommodate rapid appearance changes. The residual memory layer can extend residual learning from the spatial domain to the Spatio-temporal domain via ConvLSTM to obtain long-term memory of the target appearance. To avoid model degradation due to sample imbalance, we introduce a weighted gradient harmonized loss to improve the discrimination of the tracker. Then, response scores can be served as a basis of the adaptive learning strategy to ensure the reliability of memory updates. The proposed method performs favorably and has been extensively validated on six benchmark datasets, including OTB-50/100, TC-128, UAV-123, and VOT-2016/2018 against several advanced methods.

Price forecasting is at the center of decision making in electricity markets. Much research has been done in forecasting energy prices for a single market while little research has been reported on forecasting price difference between markets, which presents higher volatility and yet plays a critical role in applications such as virtual trading. To this end, this paper takes the first attempt at it and employs novel deep learning architecture with Bidirectional Long-Short Term Memory (LSTM) units and Sequence-to-Sequence (Seq2Seq) architecture to forecast nodal price difference between day-ahead and real-time markets. In addition to value prediction, these deep learning architectures are also used to develop classification models to predict the price difference bands/ranges. The proposed methods are tested using historical PJM market data, and evaluated using Root Mean Squared Error (RMSE) and other customized performance metrics. Case studies show that both deep learning methods outperform common methods including ARIMA, XGBoost and Support Vector Regression (SVR) methods. More importantly, the deep learning methods can capture the magnitude and timing of price difference spikes. Numerical results show the Seq2Seq model performs particularly well and demonstrates generalization capability to extended forecasting lead time.

Solar energy systems (SES) and photovoltaic (PV) modules should be operated at the maximum power point (MPP) to achieve the highest efficiency in the energy generation processes. Maximum power point tracking (MPPT) applications using conventional methods may not be able to follow the global MPP (GMPP) of the PV system under changing atmospheric conditions and they could oscillate around the local MPP. In this study, a machine learning and deep learning (DL) based long short-term memory (LSTM) model is proposed as an innovative solution for MPPT. Contrary to the traditional MPPT applications using current and voltage sensors, the output resistance of the PV module estimation was made by using environmental parameters (such as temperature and radiation) and artificial intelligence algorithms in this study.The LSTM model was compared with artificial neural networks (ANN) and regression methods regarding mean square error (MSE), root mean square error(RMSE) and mean absolute error (MAE) parameters. It has been determined that the LSTM model has a better performance and could more successfully follow MPP compared to the other methods. Finally, after the comparison with the ANN method, it is proved that LSTM gives 37%, 21%, and 31% more successful MSE, RMSE, and MAE results, respectively

Coal pulverizing systems reliability can be ensured effectively by using prognostics and health management approach. A mathematical model of coal pulverizing system used for anomaly detection is hard to be constructed due to its dynamic and nonlinear high-dimensional system typically. This paper proposed the use of the Long-Short Term Memory Autoencoder model for anomaly detection of the coal pulverizing system on a coal-fired power plant. The LSTM will solve the gradient reduction problem, and Autoencoder will improve the generalizability of the model. As a result, the proposed model can detect the anomaly successfully before the Sequent of Events occurs.

This paper presents the detection of High Impedance Fault (HIF) in solar Photovoltaic (PV) integrated power system using recurrent neural network-based Long Short-Term Memory (LSTM) approach. For study this, an IEEE 13-bus system was modeled in MATLAB/Simulink environment to integrate 300 kW solar PV systems for analysis. Initially, the three-phase current signal during non-faulty (regular operation, capacitor switching, load switching, transformer inrush current) and faulty (HIF, symmetrical and unsymmetrical fault) conditions were used for extraction of features. The signal processing technique of Discrete Wavelet Transform with db4 mother wavelet was applied to extract each phase's energy value features for training and testing the classifiers. The proposed LSTM classifier gives the overall classification accuracy of 91.21% with a success rate of 92.42 % in identifying HIF in PV integrated power network. The prediction results obtained from the proffered method are compared with other well-known classifiers of K-Nearest neighbor's network, Support vector machine, J48 based decision tree, and Naïve Bayes approach. Further, the classifier's robustness is validated by evaluating the performance indices (PI) of kappa statistic, precision, recall, and F-measure. The results obtained reveal that the proposed LSTM network significantly outperforms all PI compared to other techniques.

Due to its complexity, financial time-series forecasting is regarded as one of the most challenging problems. During the past two decades, nonlinear modeling techniques, such as artificial neural networks, were commonly employed to solve a variety of time-series problems. Recently, however, deep neural network has been found to be more efficient than those in many application domains. In this article, we propose a model based on deep neural networks that improves the forecasting of stock prices. We investigate the impact of combining deep learning techniques with multiresolution analysis to improve the forecasting accuracy. Our proposed model is based on an empirical wavelet transform which is shown to outperform traditional stationary wavelet transform in capturing price fluctuations at different time scales. The proposed model is demonstrated to be substantially more effective than other models when evaluated on the S&P500 stock index and Mackey-Glass time series.

Enhancers are short DNA segments that play a key role in biological processes, such as accelerating transcription of target genes. Since the enhancer resides anywhere in a genome sequence, it is difficult to precisely identify enhancers. We presented a bi-directional long-short term memory (Bi-LSTM) and attention-based deep learning method (Enhancer-LSTMAtt) for enhancer recognition. Enhancer-LSTMAtt is an end-to-end deep learning model that consists mainly of deep residual neural network, Bi-LSTM, and feed-forward attention. We extensively compared the Enhancer-LSTMAtt with 19 state-of-the-art methods by 5-fold cross validation, 10-fold cross validation and independent test. Enhancer-LSTMAtt achieved competitive performances, especially in the independent test. We realized Enhancer-LSTMAtt into a user-friendly web application. Enhancer-LSTMAtt is applicable not only to recognizing enhancers, but also to distinguishing strong enhancer from weak enhancers. Enhancer-LSTMAtt is believed to become a promising tool for identifying enhancers.

In order to maintain the stable operation of power systems, quick and accurate transient stability assessment (TSA) after the fault clearance is important. Machine learning methods have been widely used in the transient stability analysis of power systems. However, how to make good use of time series data of PMUs and effectively balance the contradiction between rapidity and accuracy brings new challenges to TSA. To address this problem, we propose an anti-jitter dynamic evaluation method based on long-short term memory (LSTM) network. In this model, the trajectory cluster characteristics of generators power angles after fault clearance are taken as inputs, and an improved LSTM is used to learn the nonlinear mapping relationship between the input characteristics and the transient stability. Meanwhile, by the use of sliding time windows and anti-jitter mechanism, a hierarchical real-time prediction framework is constructed to effectively utilize the time series data of PMUs. The case studies on two systems indicate that the proposed method has superior evaluation accuracy and general performance. In addition, the proposed method can effectively evaluate the stability margin or instability degree of samples, which provides reliable reference information for emergency control.

Mención Internacional en el título de doctor The improvements of the last two decades in data modeling and computing have lead to new biometric modalities. The Electrocardiogram (ECG) modality is part of them, and has been mainly researched by using public databases related to medical training. Despite of being useful for initial approaches, they are not representative of a real biometric environment. In addition, publishing and creating a new database is none trivial due to human resources and data protection laws. The main goal of this thesis is to successfully use ECG as a biometric signal while getting closer to the real case scenario. Every experiment considers low computational calculations and transformations to help in potential portability. The core experiments in this work come from a private database with different positions, time and heart rate scenarios. An initial segmentation evaluation is achieved with the help of fiducial point detection which determines the QRS selection as the input data for all the experiments. The approach of training a model per user (open-set) is tested with different machine learning algorithms, only getting an acceptable result with Gaussian Mixture Models (GMM). However, the concept of training all users in one model (closed-set) shows more potential with Linear Discriminant Analysis (LDA), whose results were improved in 40%. The results with LDA are also tested as a multi-modality technique, decreasing the Equal Error Rate (EER) of fingerprint verification in up to 70.64% with score fusion, and reaching 0% in Protection Attack Detection (PAD). The Multilayer Perceptron (MLP) algorithm enhances these results in verification while applying the first differentiation to the signal. The network optimization is achieved with EER as an observation metric, and improves the results of LDA in 22% for the worst case scenario, and decreases the EER to 0% in the best case. Complexity is added creating a Convolutional Neural Network (CNN) and Long-Short Term Memory (LSTM) based network, BioECG. The tuning process is achieved without extra feature transformation and is evaluated through accuracy, aiming for good identification. The inclusion of a second day of enrollment in improves results from MLP, reaching the overall lowest results of 0.009%–1.352% in EER. Throughout the use of good quality signals, position changes did not noticeably impact the verification. In addition, collecting data in a different day or in a different hour did not clearly affect the performance. Moreover, modifying the verification process based on attempts, improves the overall results, up to reach a 0% EER when applying BioECG. Finally, to get closer to a real scenario, a smartband prototype is used to collect new databases. A private database with limited scenarios but controlled data, and another local database with a wider range of scenarios and days, and with a more relaxed use of the device. Applying the concepts of first differentiation and MLP, these signals required the Stationary Wavelet Transform (SWT) and new fiducial point detection to improve their results. The first database gave subtle chances of being used in identification with up to 78.2% accuracy, but the latter was completely discarded for this purpose. These realistic experiments show the impact of a low fidelity sensor, even considering the same modifications in previous successful experiments with better quality data, reaching up to 13.530% EER. In the second database, results reach a range of 0.068%–31.669% EER. This type of sensor is affected by heart rate changes, but also by position variations, given its sensitivity to movement. Las mejoras en modelado de datos y computación de las últimas dos décadas, han llevado a la creación de nuevas modalidades biométricas. La modalidad de electrocardiograma (ECG) es una de ellas, la cual se ha investigado usando bases de datos públicas que fueron creadas para entrenamiento de profesional médico. Aunque estos datos han sido útiles para los estados iniciales de la modalidad, no son representativos de un entorno biométrico real. Además, publicar y crear bases de datos nuevas son problemas no triviales debido a los recursos humanos y las leyes de protección de datos. El principal objetivo de esta tesis es usar exitosamente datos de ECG como señales biométricas a la vez que nos acercamos a un escenario realista. Cada experimento considera cálculos y transformadas de bajo coste computacional para ayudar en su potencial uso en aparatos móviles. Los principales experimentos de este trabajo se producen con una base de datos privada con diferentes escenarios en términos de postura, tiempo y frecuencia cardíaca. Con ella se evalúan las diferentes seleccións del complejo QRS mediante detección de puntos fiduciales, lo cual servirá como datos de entrada para el resto de experimentos. El enfoque de entrenar un modelo por usuario (open-set) se prueba con diferentes algoritmos de aprendizaje máquina (machine learning), obteniendo resultados aceptables únicamente mediante el uso de modelos de mezcla de Gaussianas (Gaussian Mixture Models, GMM). Sin embargo, el concepto de entrenar un modelo con todos los usuarios (closed-set) demuestra mayor potencial con Linear Discriminant Analysis (Análisis de Discriminante Lineal, LDA), cuyos resultados mejoran en un 40%. Los resultados de LDA también se utilizan como técnica multi-modal, disminuyendo la Equal Error Rate (Tasa de Igual Error, EER) de la verificación mediante huella en hasta un 70.64% con fusión de puntuación, y llegando a un sistema con un 0% de EER en Detección de Ataques de Presentación (Presentation Attack Detection, PAD). El algoritmo de Perceptrón Multicapa (Multilayer Perceptron, MLP) mejora los resultados previos en verificación aplicando la primera derivada a la señal. La optimización de la red se consigue en base a su EER, mejora la de LDA en hasta un 22% en el peor caso, y la lleva hasta un 0% en el mejor caso. Se añade complejidad creando una red neural convolucional (Convolutional Neural Network, CNN) con una red de memoria a largo-corto plazo (Long-Short Term Memory, LSTM), llamada BioECG. El proceso de ajuste de hiperparámetros se lleva acabo sin transformaciones y se evalúa observando la accuracy (precisión), para mejorar la identificación. Sin embargo, incluir un segundo día de registro (enrollment) con BioECG, estos resultados mejoran hasta un 74% para el peor caso, llegando a los resultados más bajos hasta el momento con 0.009%–1.352% en la EER. Durante el uso de señales de buena calidad, los cambios de postura no afectaron notablemente a la verificación. Además, adquirir los datos en días u horas diferentes tampoco afectó claramente a los resultados. Asimismo, modificar el proceso de verificación en base a intentos también produce mejoría en todos los resultados, hasta el punto de llegar a un 0% de EER cuando se aplica BioECG. Finalmente, para acercarnos al caso más realista, se usa un prototipo de pulsera para capturar nuevas bases de datos. Una base de datos privada con escenarios limitados pero datos más controlados, y otra base de datos local con más espectro de escenarios y días y un uso del dispositivo más relajado. Para estos datos se aplican los conceptos de primera diferenciación en MLP, cuyas señales requieren la Transformada de Wavelet Estacionaria (Stationary Wavelet Transform, SWT) y un detector de puntos fiduciales para mejorar los resultados. La primera base de datos da opciones a ser usada para identificación con un máximo de precisión del 78.2%, pero la segunda se descartó completamente para este propósito. Estos experimentos más realistas demuestran el impact de tener un sensor de baja fidelidad, incluso considerando las mismas modificaciones que previamente tuvieron buenos resultados en datos mejores, llegando a un 13.530% de EER. En la segunda base de datos, los resultados llegan a un rango de 0.068%–31.669% en EER. Este tipo de sensor se ve afectado por las variaciones de frecuencia cardíaca, pero también por el cambio de posición, dado que es más sensible al movimiento. Programa de Doctorado en Ingeniería Eléctrica, Electrónica y Automática por la Universidad Carlos III de Madrid Presidente: Cristina Conde Vilda.- Secretario: Mariano López García.- Vocal: Young-Bin Known

A novel ppb-level biomedical sensor is developed to analyze breath samples for continuous monitoring of diseases. The setup is very compact, consisting of a distributed feedback quantum cascade laser (DFB-QCL) and a single-pass absorption cell. To make the sensor more compact and functional, a deep neural network (DNN) model is utilized for predicting gas concentrations. In order to evaluate the performance of the sensor, N2O is used as the target molecule. A minimum detection limit of 500 ppb is achieved in a single-pass absorption cell configuration. The model is trained on multiple N2O/CO2 absorption lines (instead of an isolated line) with concentrations between 0 to 500 ppm generated using the HITRAN database. The trained model is tested on measured spectra and compared to a non-linear least squares fitting algorithm. The coefficients of determination (R2) were found to be 0.997 and 0.981 for the predictions of N2O concentrations in the N2O/N2 gas mixture and the breath air, respectively. The accuracies of 2.5% and 2.9% were achieved by the sensor for both cases.

Facial movements are crucial for human interaction because they provide relevant information on verbal and non-verbal communication and social interactions. From a clinical point of view, the analysis of facial movements is important for diagnosis, follow-up, drug therapy, and surgical treatment. Current methods of assessing facial palsy are either (i) objective but inaccurate, (ii) subjective and, thus, depending on the clinician’s level of experience, or (iii) based on static data. To address the aforementioned problems, we implemented a deep learning algorithm to assess facial movements during smiling. Such a model was trained on a dataset that contains healthy smiles only following an anomaly detection strategy. Generally speaking, the degree of anomaly is computed by comparing the model’s suggested healthy smile with the person’s actual smile. The experimentation showed that the model successfully computed a high degree of anomaly when assessing the patients’ smiles. Furthermore, a graphical user interface was developed to test its practical usage in a clinical routine. In conclusion, we present a deep learning model, implemented on open-source software, designed to help clinicians to assess facial movements.

The gradual transition from a traditional transportation system to an intelligent transportation system (ITS) has paved the way to preserve green environments in metro cities. Moreover, electric vehicles (EVs) seem to be beneficial choices for traveling purposes due to their low charging costs, low energy consumption, and reduced greenhouse gas emission. However, a single failure in an EV’s intrinsic components can worsen travel experiences due to poor charging infrastructure. As a result, we propose a deep learning and blockchain-based EV fault detection framework to identify various types of faults, such as air tire pressure, temperature, and battery faults in vehicles. Furthermore, we employed a 5G wireless network with an interplanetary file system (IPFS) protocol to execute the fault detection data transactions with high scalability and reliability for EVs. Initially, we utilized a convolutional neural network (CNN) and a long-short term memory (LSTM) model to deal with air tire pressure fault, anomaly detection for temperature fault, and battery fault detection for EVs to predict the presence of faulty data, which ensure safer journeys for users. Furthermore, the incorporated IPFS and blockchain network ensure highly secure, cost-efficient, and reliable EV fault detection. Finally, the performance evaluation for EV fault detection has been simulated, considering several performance metrics, such as accuracy, loss, and the state-of-health (SoH) prediction curve for various types of identified faults. The simulation results of EV fault detection have been estimated at an accuracy of 70% for air tire pressure fault, anomaly detection of the temperature fault, and battery fault detection, with R2 scores of 0.874 and 0.9375.

In this study, the hourly directions of eight banking stocks in Borsa Istanbul were predicted using linear-based, deep-learning (LSTM) and ensemble learning (LightGBM) models. These models were trained with four different feature sets and their performances were evaluated in terms of accuracy and F-measure metrics. While the first experiments directly used the own stock features as the model inputs, the second experiments utilized reduced stock features through Variational AutoEncoders (VAE). In the last experiments, in order to grasp the effects of the other banking stocks on individual stock performance, the features belonging to other stocks were also given as inputs to our models. While combining other stock features was done for both own (named as allstock_own) and VAE-reduced (named as allstock_VAE) stock features, the expanded dimensions of the feature sets were reduced by Recursive Feature Elimination. As the highest success rate increased up to 0.685 with allstock_own and LSTM with attention model, the combination of allstock_VAE and LSTM with the attention model obtained an accuracy rate of 0.675. Although the classification results achieved with both feature types was close, allstock_VAE achieved these results using nearly 16.67% less features compared to allstock_own. When all experimental results were examined, it was found out that the models trained with allstock_own and allstock_VAE achieved higher accuracy rates than those using individual stock features. It was also concluded that the results obtained with the VAE-reduced stock features were similar to those obtained by own stock features.

The COVID-19 disease for Novel coronavirus (SARS-CoV-2) has turned out to be a global pandemic. The high transmission rate of this pathogenic virus demands an early prediction and proper identification for the subsequent treatment. However, polymorphic nature of this virus allows it to adapt and sustain in different kinds of environment which makes it difficult to predict. On the other hand, there are other pathogens like SARS-CoV-1, MERS-CoV, Ebola, Dengue, and Influenza as well, so that a predictor is highly required to distinguish them with the use of their genomic information. To mitigate this problem, in this work COVID-DeepPredictor is proposed on the framework of deep learning to identify an unknown sequence of these pathogens. COVID-DeepPredictor uses Long Short Term Memory as Recurrent Neural Network for the underlying prediction with an alignment-free technique. In this regard, k-mer technique is applied to create Bag-of-Descriptors (BoDs) in order to generate Bag-of-Unique-Descriptors (BoUDs) as vocabulary and subsequently embedded representation is prepared for the given virus sequences. This predictor is not only validated for the dataset using K-fold cross-validation but also for unseen test datasets of SARS-CoV-2 sequences and sequences from other viruses as well. To verify the efficacy of COVID-DeepPredictor, it has been compared with other state-of-the-art prediction techniques based on Linear Discriminant Analysis, Random Forests, and Gradient Boosting Method. COVID-DeepPredictor achieves 100% prediction accuracy on validation dataset while on test datasets, the accuracy ranges from 99.51 to 99.94%. It shows superior results over other prediction techniques as well. In addition to this, accuracy and runtime of COVID-DeepPredictor are considered simultaneously to determine the value of k in k-mer, a comparative study among k values in k-mer, Bag-of-Descriptors (BoDs), and Bag-of-Unique-Descriptors (BoUDs) and a comparison between COVID-DeepPredictor and Nucleotide BLAST have also been performed. The code, training, and test datasets used for COVID-DeepPredictor are available at http://www.nitttrkol.ac.in/indrajit/projects/COVID-DeepPredictor/.

Load models have evolved from simple ZIP model to composite model that incorporates the transient dynamics of motor loads. This research utilizes the latest trend on Machine Learning and builds reliable and accurate composite load model. A composite load model is a combination of static (ZIP) model paralleled with a dynamic model. The dynamic model, recommended by Western Electricity Coordinating Council (WECC), is an induction motor representation. In this research, a dual cage induction motor with 20 parameters pertaining to its dynamic behavior, starting behavior, and per unit calculations is used as a dynamic model. For machine learning algorithms, a large amount of data is required. The required PMU field data and the corresponding system models are considered Critical Energy Infrastructure Information (CEII) and its access is limited. The next best option for the required amount of data is from a simulating environment like PSSE. The IEEE 118 bus system is used as a test setup in PSSE and dynamic simulations generate the required data samples. Each of the samples contains data on Bus Voltage, Bus Current, and Bus Frequency with corresponding induction motor parameters as target variables. It was determined that the Artificial Neural Network (ANN) with multivariate input to single parameter output approach worked best. Recurrent Neural Network (RNN) is also experimented side by side to see if an additional set of information of timestamps would help the model prediction. Moreover, a different definition of a dynamic model with a transfer function-based load is also studied. Here, the dynamic model is defined as a mathematical representation of the relation between bus voltage, bus frequency, and active/reactive power flowing in the bus. With this form of load representation, Long-Short Term Memory (LSTM), a variation of RNN, performed better than the concurrent algorithms like Support Vector Regression (SVR). The result of this study is a load model consisting of parameters defining the load at load bus whose predictions are compared against simulated parameters to examine their validity for use in contingency analysis. Master of Science Independent system Operators (ISO) and Distribution system operators (DSO) have a responsibility to provide uninterrupted power supply to consumers. That along with the longing to keep operating cost minimum, engineers and planners study the system beforehand and seek to find the optimum capacity for each of the power system elements like generators, transformers, transmission lines, etc. Then they test the overall system using power system models, which are mathematical representation of the real components, to verify the stability and strength of the system. However, the verification is only as good as the system models that are used. As most of the power systems components are controlled by the operators themselves, it is easy to develop a model from their perspective. The load is the only component controlled by consumers. Hence, the necessity of better load models. Several studies have been made on static load modeling and the performance is on par with real behavior. But dynamic loading, which is a load behavior dependent on time, is rather difficult to model. Some attempts on dynamic load modeling can be found already. Physical component-based and mathematical transfer function based dynamic models are quite widely used for the study. These load structures are largely accepted as a good representation of the systems dynamic behavior. With a load structure in hand, the next task is estimating their parameters. In this research, we tested out some new machine learning methods to accurately estimate the parameters. Thousands of simulated data are used to train machine learning models. After training, we validated the models on some other unseen data. This study finally goes on to recommend better methods to load modeling.

Physicians manually interpret an electrocardiogram (ECG) signal morphology in routine clinical practice. This activity is a monotonous and abstract task that relies on the experience of understanding ECG waveform meaning, including P-wave, QRS-complex, and T-wave. Such a manual process depends on signal quality and the number of leads. ECG signal classification based on deep learning (DL) has produced an automatic interpretation; however, the proposed method is used for specific abnormality conditions. When the ECG signal morphology change to other abnormalities, it cannot proceed automatically. To generalize the automatic interpretation, we aim to delineate ECG waveform. However, the output of delineation process only ECG waveform duration classes for P-wave, QRS-complex, and T-wave. It should be combined with a medical knowledge rule to produce the abnormality interpretation. The proposed model is applied for atrial fibrillation (AF) identification. This study meets the AF criteria with RR irregularities and the absence of P-waves in essential oscillations for even more accurate identification. The QT database by Physionet is utilized for developing the delineation model, and it validates with The Lobachevsky University Database. The results show that our delineation model works properly, with 98.91% sensitivity, 99.01% precision, 99.79% specificity, 99.79% accuracy, and a 98.96% F1 score. We use about 4058 normal sinus rhythm records and 1804 AF records from the experiment to identify AF conditions that are taken from three datasets. The comprehensive testing has produced higher negative predictive value and positive predictive value. This means that the proposed model can identify AF conditions from ECG signal delineation. Our approach can considerably contribute to AF diagnosis with these results.

River flow modeling plays a crucial role in water resource management and ensuring its sustainability. Therefore, in recent years, in addition to the prediction of hydrological processes through modeling, applicable and highly reliable methods have also been used to analyze the sustainability of water resources. Artificial neural networks and deep learning-based hybrid models have been used by scientists in river flow predictions. Therefore, in this study, we propose a hybrid approach, integrating long-short-term memory (LSTM) networks and a genetic algorithm (GA) for streamflow forecasting. The performance of the hybrid model and the benchmark model was taken into account using daily flow data. For this purpose, the daily river flow time series of the Beyderesi-Kılayak flow measurement station (FMS) from September 2000 to June 2019 and the data from Yazıköy from December 2000 to June 2018 were used for flow measurements on the Euphrates River in Turkey. To validate the performance of the model, the first 80% of the data were used for training, and the remaining 20% were used for the testing of the two FMSs. Statistical methods such as linear regression was used during the comparison process to assess the proposed method’s performance and to demonstrate its superior predictive ability. The estimation results of the models were evaluated with RMSE, MAE, MAPE, STD and R2 statistical metrics. The comparison of daily streamflow predictions results revealed that the LSTM-GA model provided promising accuracy results and mainly presented higher performance than the benchmark model and the linear regression model.

This paper presents a novel method for attitude estimation of an object in 3D space by incremental learning of the Long-Short Term Memory (LSTM) network. Gyroscope, accelerometer, and magnetometer are few widely used sensors in attitude estimation applications. Traditionally, multi-sensor fusion methods such as the Extended Kalman Filter and Complementary Filter are employed to fuse the measurements from these sensors. However, these methods exhibit limitations in accounting for the uncertainty, unpredictability, and dynamic nature of the motion in real-world situations. In this paper, the inertial sensors data are fed to the LSTM network which are then updated incrementally to incorporate the dynamic changes in motion occurring in the run time. The robustness and efficiency of the proposed framework is demonstrated on the dataset collected from a commercially available inertial measurement unit. The proposed framework offers a significant improvement in the results compared to the traditional method, even in the case of a highly dynamic environment. The LSTM framework-based attitude estimation approach can be deployed on a standard AI-supported processing module for real-time applications.

This article describes how the human activity recognition in videos is a very attractive topic among researchers due to vast possible applications. This article considers the analysis of behaviors and activities in videos obtained with low-cost RGB cameras. To do this, a system is developed where a video is input, and produces as output the possible activities happening in the video. This information could be used in many applications such as video surveillance, disabled person assistance, as a home assistant, employee monitoring, etc. The developed system makes use of the successful techniques of Deep Learning. In particular, convolutional neural networks are used to detect features in the video images, meanwhile Recurrent Neural Networks are used to analyze these features and predict the possible activity in the video. This work has been funded by the Spanish Government TIN2016-76515-R grant for the COMBAHO project, supported with Feder funds.

The application of artificial neural network to load forecasting can overcome the problem of dynamic load change, and its ability to adapt to nonlinear relationships makes the prediction result satisfactory. This paper firstly reviews and introduces the concepts and basic principles of load prediction, discusses various methods for load forecasting, and then selects artificial neural network to establish a predictive model. In this paper, the European electric load is predicted with a BP neural network. From the prediction results, it is feasible to use BP neural network for load forecasting, and its accuracy can meet the needs of real-life engineering work. However, BP neural networks have the problem of slow convergence and easily falling into local minimum points. Therefore, this paper also uses three other neural networks for load forecasting, which are Radial Basis Network (RBF), Elman Network, and Long-Short Term Memory Network (LSTM). In the experiment, the four neural networks achieved expected prediction results, and the LSTM network had the best prediction effect. Scientific discussions are offered.

Background Clinical entity recognition as a fundamental task of clinical text processing has been attracted a great deal of attention during the last decade. However, most studies focus on clinical text in English rather than other languages. Recently, a few researchers have began to study entity recognition in Chinese clinical text. Methods In this paper, a novel deep neural network, called attention-based CNN-LSTM-CRF, is proposed to recognize entities in Chinese clinical text. Attention-based CNN-LSTM-CRF is an extension of LSTM-CRF by introducing a CNN (convolutional neural network) layer after the input layer to capture local context information of words of interest and an attention layer before the CRF layer to select relevant words in the same sentence. Results In order to evaluate the proposed method, we compare it with other two currently popular methods, CRF (conditional random field) and LSTM-CRF, on two benchmark datasets. One of the datasets is publically available and only contains contiguous clinical entities, and the other one is constructed by us and contains contiguous and discontiguous clinical entities. Experimental results show that attention-based CNN-LSTM-CRF outperforms CRF and LSTM-CRF. Conclusions CNN and attention mechanism are individually beneficial to LSTM-CRF-based Chinese clinical entity recognition system, no matter whether contiguous clinical entities are considered. The conribution of attention mechanism is greater than CNN.

Mental fatigue (MF) monitoring is essential for eliminating accidents in high-risk tasks and providing better productivity management in daily work tasks involving human operators. Previous works only built MF monitoring systems with either high time resolution or high granularity level. We proposed a physiological-based approach to estimate MF Level every 2 seconds in a regression manner, a system that realized both high time resolution and high granularity level. The approach consists of an accurate MF level assessment method using the alignment score within a modified N-back task that is free from the lure effect, a long short-term memory (LSTM) deep learning framework, and a performance-reliability validation process. We used multiple physiological signals including ECG, respiration, and pupil diameter. As a result, we provided feasible estimating performance not worse than existing studies while realizing both high time resolution and high level of granularity. The interpretation analysis utilized accumulated local effects (ALE) to illustrate how black-box models make estimations and to improve the reliability of this approach.

A smart home provides a facilitated environment for the detection of human activity with appropriate Deep Learning algorithms to manipulate data collected from numerous sensors attached to various smart things in a smart home environment. Human activities comprise expected and unexpected behavior events, therefore, detecting these events consisting of mutual dependent activities poses a key challenge in the activities detection paradigm. Besides, the battery-powered sensor ubiquitously and extensively monitors activities, disputes, and sensor energy depletion. Therefore, to address these challenges, we propose an Energy and Event Aware-Sensor Duty Cycling scheme. The proposed model predicts the future expected event using the Bi-Directional Long-Short Term Memory model and allocates Predictive Sensors to the predicted event. To detect the unexpected events, the proposed model localizes a Monitor Sensor within a cluster of Hibernate Sensors using the Jaccard Similarity Index. Finally, we optimize the performance of our proposed scheme by employing the Q-Learning algorithm to track the missed or undetected events. The simulation is executed against the conventional Machine Learning algorithms for the sensor duty cycle, scheduling to reduce the sensor energy consumption and improve the activity detection accuracy. The experimental evaluation of our proposed scheme shows significant improvement in activity detection accuracy from 94.12% to 96.12%. Besides, the effective rotation of the Monitor Sensor significantly improves the energy consumption of each sensor with the entire network lifetime.

Accompanied by the development of new energy resources, lithium-ion batteries have been used widely in various fields. Due to the significant influence of system performance, much attention has been paid to the accurate estimation and prediction about health status of lithium-ion batteries. In a battery pack, the structure connection causes sophisticated interaction between cells, or between the cells and the pack. Therefore, the degradation of any cell is the result of the deterioration of conjoint cells, and a rapid degradation speed for any individual cell can lead to the accelerated degradation of others beyond expectation, which is one of the primary reasons why the State of Health and life cannot be calculated precisely. To solve this problem, a novel method based on integrated state information from cells has been proposed to estimate status of packs, considering about the degradation effect that cells contribute to the corresponding pack. Using this method, the interactive relationship was described in the form of a neural network in order to mine the effect from the inter-degradation between cells. It was proven that the novel method had better performance than a method based only on the degradation indicators from battery packs.

Smart WiFi thermostats have moved well beyond the function they were originally designed for, namely, controlling heating and cooling comfort in buildings. They are now also learning from occupant behaviors and permit occupants to control their comfort remotely. This research seeks to go beyond this state of the art by utilizing smart WiFi thermostat data in residences to develop dynamic predictive models for room temperature and cooling/heating demand. These models can then be used to estimate the energy savings from new thermostat temperature schedules and estimate peak load reduction achievable from maintaining a residence in a minimum thermal comfort condition. Back Propagation Neural Network (BPNN), Long-Short Term Memory (LSTM), and Encoder-Decoder LSTM dynamic models are explored. Results demonstrate that LSTM outperforms BPNN and Encoder-Decoder LSTM approach, yielding and a MAE error of 0.5 &deg, C, equal to the resolution error of the measured temperature. Additionally, the models developed are shown to be highly accurate in predicting savings from aggressive thermostat set point schedules, yielding deep reduction of up to 14.3% for heating and cooling, as well as significant energy reduction from curtailed thermal comfort in response to a high demand event.

The escalator is one of the most popular travel methods in public places, and the safe working of the escalator is significant. Accurately predicting the escalator failure time can provide scientific guidance for maintenance to avoid accidents. However, failure data have features of short length, non-uniform sampling, and random interference, which makes the data modeling difficult. Therefore, a strategy that combines data quality enhancement with deep neural networks is proposed for escalator failure time prediction in this paper. First, a comprehensive selection indicator (CSI) that can describe the stationarity and complexity of time series is established to select inherently excellent failure sequences. According to the CSI, failure sequences with high stationarity and low complexity are selected as the referenced sequences to enhance the quality of other failure sequences by using dynamic time warping preprocessing. Secondly, a deep neural network combining the advantages of a convolutional neural network and long short-term memory is built to train and predict quality-enhanced failure sequences. Finally, the failure-recall record of six escalators used for 6 years is analyzed by using the proposed method as a case study, and the results show that the proposed method can reduce the average prediction error of failure time to less than one month.

Gürültülü veri, özellikle kullanıcı tarafından oluşturulan içerik (örn. çevrimiçi yorum, tweet), üzerinde varlık ismi tanıma metnin bozuk yapısından dolayı zorlu bir görevdir. Bu kapsamda, İngilizce gibi göreli morfolojik olarak fakir diller üzerindeki araştırmalar son yıllarda hızla ilerlerken, Türkçe gibi morfolojik olarak zengin dillerde gürültülü veri üzerindeki araştırmalar geri kalmıştır. Bu çoğunlukla Türkçe dilinin morfolojik olarak zengin ve sondan eklemeli olması ile az miktarda etiketli veriye sahip olması nedeniyledir. Türkçe'de varolan araştırmalar hem gürültü hem de resmi (örn. haber metni) veri üzerinde çoğunlukla hala el yapımı öznitelikler ve/veya alana-özgü harici kaynaklardan (isim listeleri) faydalanmaktadır. Bu tezdeyse, el yapımı öznitelikler ve/veya alana-özgü harici kaynaklar kullanmayan yapay sinir ağlarının etkileri incelenmektedir. Öyle ki, önerilen model farklı morfolojik olarak dillerde ve farklı alanlarda da kullanılabilsin. Bununla birlikte, farklı kelime ve kelime-altı (örn. morfem ve karakter n-gram seviyesinde) embedding teknikleriyle de deneyler gerçekleştirdik ve morfolojik olarak zengin diller için kelime-altı embedding'lerin sözdizimi ve anlamsal açıdan daha iyi kelime temsili sunduğunu savunuyoruz. Bu amaçla, temel aldığımız LSTM-CRF mimarisinin uzantısı olan bir transfer öğrenme modeli önermekteyiz. Söz konusu model, resmi veriden gürültülü veriye bilgi aktarımı amacıyla aynı anda iki farklı veri kümesi üzerinde eğitilmekte olup; morfem, karakter n-gram ve ortografik embedding'lerden faydalanır. Sonuç olarak, Türkçe gürültülü veri kümesi üzerinde %65.72 ve İngilizce WNUT'17 veri kümesinde %41.97 F1 puanı elde ettik. Named entity recognition (NER) on noisy data, specifically user-generated content (e.g. online reviews, tweets) is a challenging task because of the presence of ill-formed text. In this regard, while studies on morphologically-poor languages such as English has been rapidly advancing in recent years, studies on morphologically-rich languages such as Turkish has fallen behind for noisy data. This is mostly due to Turkish being an agglutinative language, having a rich morphology and also having scarce annotated data. Existing studies on Turkish both for noisy and formal (e.g. news text) data still make use of hand-crafted features and/or external domain-specific resources (e.g. gazetteers). In this thesis, we investigate the effects of neural architectures without the help of any external domain-specific resources and/or manually-constructed features. So that the proposed model can also be used for different morphologically-rich languages and for different domains. Moreover, we also experimented with different word and sub-word level (e.g. morpheme, character or character n-gram level) embedding techniques and we argue that sub-word level embeddings provide better word representations for morphologically-rich languages syntactically and semantically. For this purpose, we propose a transfer learning model that is an extension of a baseline, bidirectional LSTM-CRF architecture. The model is trained on two different datasets simultaneously for the purpose of transfer learning from formal to noisy data and it exploits morpheme-level, character n-gram level and orthographic character-level embeddings as its feature set. Consequently, we have obtained an F1 score of 65.72% on Turkish tweet dataset and 41.97% on English WNUT'17 dataset. 127

A biomechanical understanding of gait stability is needed to reduce falling risk. As a typical parameter, the COM-COP (center of mass&ndash, center of pressure) inclination angle (IA) could provide valuable insight into postural control and balance recovery ability. In this study, an artificial neural network (ANN) model was developed to estimate COM-COP IA based on signals using an inertial sensor. Also, we evaluated how different types of ANN and the cutoff frequency of the low-pass filter applied to input signals could affect the accuracy of the model. An inertial measurement unit (IMU) including an accelerometer, gyroscope, and magnetometer sensors was fabricated as a prototype. The COM-COP IA was calculated using a 3D motion analysis system including force plates. In order to predict the COM-COP IA, a feed-forward ANN and long-short term memory (LSTM) network was developed. As a result, the feed-forward ANN showed a relative root-mean-square error (rRMSE) of 15% while the LSTM showed an improved accuracy of 9% rRMSE. Additionally, the LSTM displayed a stable accuracy regardless of the cutoff frequency of the filter applied to the input signals. This study showed that estimating the COM-COP IA was possible with a cheap inertial sensor system. Furthermore, the neural network models in this study can be implemented in systems to monitor the balancing ability of the elderly or patients with impaired balancing ability.

This study applied a GR4J model in the Xiangjiang and Qujiang River basins for rainfall-runoff simulation. Four recurrent neural networks (RNNs)&mdash, the Elman recurrent neural network (ERNN), echo state network (ESN), nonlinear autoregressive exogenous inputs neural network (NARX), and long short-term memory (LSTM) network&mdash, were applied in predicting discharges. The performances of models were compared and assessed, and the best two RNNs were selected and integrated with the lumped hydrological model GR4J to forecast the discharges, meanwhile, uncertainties of the simulated discharges were estimated. The generalized likelihood uncertainty estimation method was applied to quantify the uncertainties. The results show that the LSTM and NARX better captured the time-series dynamics than the other RNNs. The hybrid models improved the prediction of high, median, and low flows, particularly in reducing the bias of underestimation of high flows in the Xiangjiang River basin. The hybrid models reduced the uncertainty intervals by more than 50% for median and low flows, and increased the cover ratios for observations. The integration of a hydrological model with a recurrent neural network considering long-term dependencies is recommended in discharge forecasting.

The purpose of this study was the development and experimental validation of a novel portable tactile braille reading device. This design aimed to address the lack of quality braille reading material in South African schools dedicated to the blind and visually impaired. The design was divided into four subsystems: the actuation of the refreshable braille display, the mechanical design of the hand mounted device, the Optical Character Recognition (OCR) software and the electronic control system. The objectives of the study included the design of the electrical, mechanical and software subsystems of the device with emphasis on miniaturisation and a low-cost design, as well as the experimental validation of the haptic feedback and OCR subsystems. The system validation focussed on the experimental analysis of the OCR software design and the Dielectric Elastomer Actuators (DEAs) utilised in the refreshable display. The performance of low-cost elastomer and electrode materials were experimentally assessed during the design of the miniature DEAs. The ideal material combination for the proposed application was identified as a VHB4910 acrylic film lined with MG Chemicals 846 carbon grease. The largest vertical deflections of the diaphragm-type DEA were attained with a stipple-based electrode application, with the elastomer prestrained to 300 % initial area. The OCR program analysis indicated that multiple character recognition was more efficient and less prone to inaccuracies than the initially proposed single character recognition method. The integration of OCR software with miniature DEAs served as a novel approach to text-to-braille transcription.