Search

Multiview learning has received substantial attention over the past decade due to its powerful capacity in integrating various types of information. Conventional unsupervised multiview dimension reduction (UMDR) methods are usually conducted in an offline manner and may fail in many real-world applications, where data arrive sequentially and the data distribution changes periodically. Moreover, satisfying the requirements of high memory consumption and expensive retraining of the time cost in large-scale scenarios are difficult. To remedy these drawbacks, we propose an online UMDR (OUMDR) framework. OUMDR aims to seek a low-dimensional and informative consensus representation for streaming multiview data. View-specific weights are also learned in this article to reflect the contributions of different views to the final consensus presentation. A specific model called OUMDR-E is developed by introducing the exclusive group LASSO (EG-LASSO) to explore the intraview and interview correlations. Then, we develop an efficient iterative algorithm with limited memory and time cost requirements for optimization, where the convergence of each update is theoretically guaranteed. We evaluate the proposed approach in video-based expression recognition applications. The experimental results demonstrate the superiority of our approach in terms of both effectiveness and efficiency.

Modeling the current-voltage (I–V) characteristics of photovoltaic (PV) modules is of importance for condition monitoring and fault diagnosis of PV arrays. The practicality of the one-diode model (ODM) is relying on the behaviors of parameter extraction and translation techniques. In this paper, a simple and robust dual-iteration algorithm is presented to extract the ODM parameters using only three points under limited conditions. These points can be easily obtained from both the datasheet and measurements, making the method universal and practical. Parameters under other unknown conditions are translated from the nearest acquired condition to make the proposed method applicable for various conditions. The modeling I–V curve obtained by the proposed method is in a close agreement with the field measurement, demonstrating its effectiveness to characterize the PV modules with different technologies. Moreover, three existing models are used to benchmark the accuracy of the proposed method. The comparison further confirms that the accuracy of the ODM has been enhanced especially at low irradiance using the proposed parameter translation method. Furthermore, a distance-weighted method is presented to reliably translate the parameters, whose superiority has been validated against measurements and the other four methods, thus being an effective output assessment for monitoring and diagnosis.

In this study, a graph regularized algorithm for early expression detection (EED), called GraphEED, is proposed. EED is aimed at detecting the specified expression in the early stage of a video. Existing EED detectors fail to explicitly exploit the local geometrical structure of the data distribution, which may affect the prediction performance significantly. According to manifold learning, the data in real-world applications are likely to reside on a low-dimensional submanifold embedded in the high-dimensional ambient space. The proposed graph Laplacian consists of two parts: 1) a k-nearest neighbor graph is first constructed to encode the geometrical information under the manifold assumption and 2) the entire expressions are regarded as the must-link constraints since they all contain the complete duration information and it is shown that this can also be formulated as a graph regularization. GraphEED is to have a detection function representing these graph structures. Even with the inclusion of the graph Laplacian, the proposed GraphEED has the same computational complexity as that of the max-margin EED, which is a well-known learning-based EED, but the detection performance has been largely improved. To further make the model appropriate in large-scale applications, with the technique of online learning, the proposed GraphEED is extended to the so-called online GraphEED (OGraphEED). In OGraphEED, the buffering technique is employed to make the optimization practical by reducing the computation and storage cost. Extensive experiments on three video-based datasets have demonstrated the superiority of the proposed methods in terms of both effectiveness and efficiency.

As the core component of the photovoltaic system, the quality of solar cells determines the conversion efficiency of electric energy. Some strategies have been proposed to detect the crack of solar cells, but most of them can not detect the crack efficiently. This paper proposed a new two-stage method for microcrack detection in polycrystalline images based on contrastive learning. First, the input picture without a label is learned by SimCLR to obtain the representation of the image. In the second stage, the linear classifier is trained based on the fixed encoder and the representation. In the comparative experiment, unsupervised contrastive learning is compared with cross-entropy training and supervised contrastive learning. The experimental results show that the linear classifier trained on unsupervised representation achieves a top-1 accuracy of 78.39%, which is 7.42% higher than the supervised contrastive learning method, compared with supervised learning, the results are comparable.

In the actual operation of photovoltaic power stations, various types of shading, including leaves, bird’s droppings, and clouds, will inevitably occur. These shadings will affect the output power of the array. Among them, if a small area of shading is not removed in time, hot spots may be formed, accelerating the aging of photovoltaic modules, and not conducive to the healthy and stable operation of the overall system. However, different shading objects cover different shading areas of photovoltaic modules, which cause different effects. This paper mainly analyzes the influence of different areas of shading on the current-voltage curve of photovoltaic modules, and proposes a classification algorithm based on weighted K-nn to classify the shading area according to the current-voltage curve characteristics of the module, which is helpful for subsequent identification of shading types and troubleshooting.

With the development of energy storage technology, more and more battery energy storage stations (BESSs) are connected to the power grid to restrain fluctuation and intermittency of renewable energy generation on the stability of the power grid. However, the connection with BESS will change the topology and power flow direction of the power grid. Therefore, based on the actual geographical wiring diagram and related data, this paper describes a grid model to analyze the influence of BESS connected to the power grid on power flow and node voltage. When the voltage of each node is stable, the genetic algorithm is applied to study how to select the access position and control the power of centralized or distributed BESSs to minimize the active power loss of the power grid. It provides guidance and reference for future location planning and control strategy of BESSs in the power grid.

As a key component of human–computer intelligent interaction and many real-world applications, the real-time property of facial expression recognition is especially important. However, the recognition result of conventional video-based approaches can not be given until the entire video is finished. In this letter, we deal with early expression detection, which aims to identify the expression as early as possible before its ending. This is a relatively new and challenging problem. Max-margin early event detector (MMED) is a well-known framework, which can make early detection. However, the linearity restricts its applications. We thus introduce kernel learning to model the nonlinear structure of complex data distribution. Moreover, the model is further reformulated in an online setting to address the streaming videos. The high retraining cost and large memory requirement of MMED are thus significantly reduced. In addition, we employ AlexNet architecture to make further comparison with mid-level features. Experiments on two popular video-based expression datasets demonstrate both the effectiveness and efficiency of the proposed method.

Cloud cover and cloud motion have a large impact on solar irradiance. One of the effective ways for direct normal irradiance (DNI) prediction is to use cloud features, which has been extensively studied. A Convolutional Neural Network (CNN) has the advantage of automatic features extraction by using strong computing capabilities. In this paper, a novel 3D-CNN method is proposed by processing multiple consecutive ground-based cloud (GBC) images in order to extract cloud features including texture and temporal information. The resulting features and the DNI data are then used to establish a DNI forecasting model. The experiments are carried out to evaluate the performance of the proposed forecasting method by using the data from January 1, 2013 to December 31, 2014. The experimental results show that the proposed method coupled with the multilayer perceptron (MLP) model achieves forecast skill of 17.06% for 10-minute ahead DNI prediction.

Extensive research has been paid for event detection in the past decades. However, the timeliness requirement of real-world applications cannot be satisfied by these approaches. Early event detector is thus proposed recently to deal with this issue. Early detection aims to recognize the target as early as possible, i.e., it can detect partial events and create a monotonous function to rank them. Although important and practical, few studies have been given for early detection due to its complexity. Max-margin Early Event Detector (MMED) is a well-known approach, which achieves satisfied performance in identifying partial events. However, the MMED works in an offline manner and may fail in this era of streaming sequence. In addition, the large memory consumption and high retraining time cost of the MMED are hard to be satisfied in general platform conditions. In this paper, we introduce an online learning technique with max-margin to early event detection. The proposed model could be adapted to the changing data distribution of the streaming sequences. No historical data need to be stored. Therefore, both the memory requirement and retraining time cost are decreased significantly. We evaluate the proposed approach on three benchmark datasets with various complexities. The extensive results demonstrate both the effectiveness and efficiency of the proposed framework.

Epileptic seizure prediction is one of the most used therapeutic adjuvant strategies for drug-resistant epilepsy. Conventional methods usually adopt handcrafted features and manual parameter setting. The over-reliance on the expertise of specialists may lead to weak exploitation of features and low popularization of clinical application. This paper proposes a novel parameterless patient-specific method based on Fourier Neural Network (FNN), where the Fourier transform and backpropagation learning are synthesized to make the predictor more efficient and practical. The employment of FNN is the first attempt in the field of seizure prediction due to its automatic extraction of immanent spectra in epileptic signals. Despite the self-adaptive superiority of FNN, we introduce Convolutional Neural Network (CNN) to further improve its search capability in high-dimensional feature spaces. The study also develops a multi-layer module to estimate spectral power ratios of raw recordings, which optimizes the prediction by enhancing feature diversity. Based on these modules, this paper proposes a two-channel deep neural network: Fourier Ratio Convolutional Neural Network (FRCNN). To demonstrate the reliability of the model, we explain the mathematical meaning of hidden-layer neurons in FRCNN theoretically. This approach is evaluated on both intracranial and scalp EEG datasets. It shows that the predictor achieved a sensitivity of 91.2% and a false prediction rate (FPR) of 0.06[Formula: see text]h[Formula: see text] across intracranial subjects and a sensitivity of 85.4% and an FPR of 0.14[Formula: see text]h[Formula: see text] over scalp subjects. The results indicate that FRCNN enables the convenience of epilepsy treatments while preserving a high degree of precision. In the end, a detailed comparison with the previous methods demonstrates that FRCNN has achieved higher performance and generalization ability.

Due to the complex geometry and wide shape variations of hair, it is difficult to collect 3D hair data for deep learning applications of hair modeling. This paper introduces a geometric example-based method to generate 3D hair data. Our key idea is to generate new hair data according to a reference hair model automatically. The reference hair model is first aligned with the most widely used parametric human model. We then edit individual strand by an offset function and generate full hair with Perlin Noise. Finally, smoother results without hair-body penetration are obtained by optimizing the moving direction. Our method can be used to generate sufficient data for deep learning applications of hair modeling and human digitization with detailed hair strands. The results demonstrate the effectiveness of our method.

Nowadays the photoelectric system that relies on solar energy is widely used. Solar cells are an important part of this system, and micro-cracks on the solar cells can reduce the overall efficiency. At present micro-cracks are detected manually by examining the electroluminescence (EL) images, which costs much manpower. To achieve the automatic detection of micro-cracks, a new self-labeling method is adopted which can label the image data automatically via simultaneous representation learning and clustering in the paper. Firstly, we build a model for representation learning by selecting ResNet-50 as the base architecture. Then the datasets are labeled automatically by adopting an algorithm that can obtain the optimal transmission of the image. At last, experiments are performed to verify the validity and the robustness of our method. The results show that the newly adopted method is more accurate than the classical clustering methods in the solar cell micro-crack data.

In recent years, the photovoltaic industry has developed vigorously where fire occurs frequently at photovoltaic power plants. This paper proposes an image defogging method based on parallel convolutional neural networks and a combined detection method for smoke and flame based on improved Faster R-CNN. Firstly, fire datasets used in the photovoltaic power plant scene are collected and produced, including two categories of smoke and flame. Then, one deep learning method is used to suppress the haze interference in the photovoltaic power plant scene. Finally, we use ResNet instead of VGG16 network, and integrate feature pyramid network(FPN) into regional candidate networks(RPN) to improve the original Faster R-CNN structure. A large number of experimental results show that our improved model has better effects on joint detection of smoke and flame.

The photovoltaic (PV) cell temperature strongly affects the performance and efficiency of the entire PV module. Thus, the accurate estimation of the cell temperature plays an important role in the health monitoring and energy assessment of PV systems. This article proposes a multi-state dynamic thermal model for PV modules, considering the heat-transfer mechanisms between the module and its environments, as well as between layers. The proposed model is benchmarked against field measurements at Aalborg University, Denmark. The results demonstrate the effectiveness of the model to characterize the internal behavior of the PV module under varying weather conditions. The performance of the proposed thermal model is also compared with prior-art models, i.e., two benchmark models, a one-state thermal model and two typical empirical equation-based models. The comparison further confirms that the estimation of cell temperature using the developed model is more accurate, presenting a reliable prediction of power production for further monitoring and diagnosis.

AAOISE, the Antarctic Astronomical Observations Intelligent Support Equipment, is an autonomous control equipment serving for energy support and environment thermal preservation, which is used for astronomical science observations in the Antarctic “Dome A”. It was deployed to “Dome A” and had an unattended run until now. The AAOISE stressed on the ways to adapt to adverse circumstances of “Dome A” and to have as little influence on the environment as possible. Its shape and structure are fully qualified for transportation and thermal insulation demands. The power generation and control systems are designed to provide continuous power and heat. Its communication system can support high-reliability data transmission and communications. It offers a possibility for developing “Dome A” scientific activities and remote monitoring of the running situation of the science instruments. This paper presents a detailed description of the power generation, power control, thermal management, instrument interface, and communications systems for AAOISE.

Solar cells play a key part in photovoltaic power system. Micro-crack is one of the common defaults, which will reduce the conversion efficiency and usable lifetime of solar cells enormously. So it is essential to locate it and identify the shape to help us obtain more detailed assessment of power loss. However, the automatic detection and segmentation of micro-crack from electroluminescence (EL) images of solar cells have been a challenging task. In this work, we propose an unsupervised multiscale micro-crack segmentation scheme for multicrystalline solar cells, which takes advantages of both of the superpixellevel and the pixel-level segmentation on the EL images. Firstly, in order to avoid disturbances from the grid of the cell, we partition the EL image of a single cell into pieces. Then, a fusion of superpixel and pixel outcomes is employed to obtain the accurate pixel-level micro-crack by considering the global and local features collectively. Finally, we restore the results of each pixel on the piece to the cell corresponding location. The experimental results show that the proposed scheme is more accurate, and the shape and area of micro-crack can be evaluated particularly.

Photovoltaic power generation is an important part of the current new energy sources. With the development of photovoltaic power generation technology, the corresponding security operation technology has become a new research topic. In this paper, a multi-level fire detection platform based on multi-source heterogeneous information fusion is proposed. The platform includes two levels of detection, level 1 is smoke detection through video monitoring data. Level 2 is the fire detection of data collected by meteorological monitoring. Through the two-level detection, the fire hazard situation and fire hazard location can be accurately located.

Solar cell efficiency is one of the most concerned issues during the photovoltaic power generation. The existed micro-crack detection approaches mainly rely on manual work with the detection efficiency is low. Besides, there lacks exploration in algorithms for separating the region of damage without much workforce. In this paper, we propose a two-stage deep scheme, especially for the polycrystalline micro-crack detection method. A region of interest (ROI) proposal method based on the canny feature and the wavelet feature for micro-crack is raised. The computational efficiency and detection accuracy are greatly improved. Firstly, we split the original electroluminescent image (EL image) by its grid line and cut them into fixed-size squares. Then the ROI proposal method is applied to extract candidate boxes as the inputs of the second stage. At the second stage, a modified convolution neural network based on the candidate boxes is supervised by binary labels. Our experimental results demonstrate that our polycrystalline micro-crack detection scheme outperforms other traditional frameworks. This work is the first attempt to solve engineering problems about micro-crack by deep learning techniques in photovoltaic systems, without many high-quality labels and computing power.

Epileptic seizures occur as a result of a disorder of brain waves. In this study, we aim at a patient-specific model for predicting the risk of seizure paroxysm by extracting the in-depth features from continuous scalp electroencephalographic(EEG) signals. Unfortunately, the false warning rates of generalized regression prediction algorithms are relatively high on account of the complexity of the cerebral time series. Hence we set the seizure forecast problem as a time series classification(TSC) problem and a cutting-edge TSC algorithm:the Hierarchical Vote Collective of Transformation-based Ensembles(HIVE-COTE) is introduced to address the issue at the first time. In addition, we improve the structure of HIVE-COTE using a channel selection module. With the improved HIVE-COTE model, an oncoming seizure could be detected by a collective of machine-learning classifiers that identify the accurate pattern(inter-ictal or pre-ictal) of the cerebral signals. Based on 15 patients in CHB-MIT dataset, the proposed improved HIVE-COTE model achieves a sensitivity of 87.4%, an AUC of 0.859, which is a promising result. Our work significantly outperform a random predictor and other TSC algorithms.

This paper considers an adaptive fault-tolerant control problem for a class of uncertain strict feedback nonlinear systems, in which the actuator has an unknown drift fault and the loss of effectiveness fault. Based on the event-triggered theory, the adaptive backstepping technique, and Lyapunov theory, a novel fault-tolerant control strategy is presented. It is shown that an appropriate comprise between the control performance and the sensor data real-time transmission consumption is made, and the fault-tolerant tracking control problem of the strict feedback nonlinear system with uncertain and unknown control direction is solved. The adaptive backstepping method is introduced to compensate the actuator faults. Moreover, a new adjustable event-triggered rule is designed to determine the sampling state instants. The overall control strategy guarantees that the output signal tracks the reference signal, and all the signals of the closed-loop systems are convergent. Finally, the fan speed control system is constructed to demonstrate the validity of the proposed strategy and the application of the general systems.

The ability to predict wind speeds is very important for the security and stability of wind farms and power system operations. Wind speeds typically vary slowly over time, which makes them difficult to forecast. In this study, a hybrid nonlinear estimation approach combining Gaussian process (GP) and unscented Kalman filter (UKF) is proposed to predict dynamic changes of wind speed and improve forecasting accuracy. The proposed approach can provide both point and interval predictions for wind speed. Firstly, the GP method is established as the nonlinear transition function of a state space model, and the covariance obtained from the GP predictive model is used as the process noise. Secondly, UKF is used to solve the state space model and update the initial prediction of short-term wind speed. The proposed hybrid approach can adjust dynamically in conjunction with the distribution changes. In order to evaluate the performance of the proposed hybrid approach, the persistence model, GP model, autoregressive (AR) model, and AR integrated with Kalman filter (KF) model are used to predict the results for comparison. Taking two wind farms in China and the National Renewable Energy Laboratory (NREL) database as the experimental data, the results show that the proposed hybrid approach is suitable for wind speed predictions, and that it can increase forecasting accuracy.

Locomotion is one of the most important problems in virtual reality. Real walking experience is the key to immersively explore the virtual world. Several strategies have been proposed to solve the problem, but most are not suitable to solve the locomotion problem in Room-Scale VR. The omnidirectional treadmill is an effective way to provide a natural walking experience within the Room-Scale VR. This paper proposes a novel omnidirectional treadmill named HEX-CORE-PROTOTYPE (HCP). The principle of synthesis and decomposition of velocity is applied to form an omnidirectional velocity field. Our system could provide a full degree of freedom and real walking experience in place. Compared to the current best system, the height of HCP is only 40% of it. The application shows the effectiveness of our system to solve the locomotion problem in Room-Scale VR.

The basis of natural interaction in VR is the exact input of human motions, which includes the locomotion, body motion, facial expressions, etc. We define the input of all such human motions as the Omnidirectional Motion Input(OMI). Traditional motion input systems are designed for a specific motion. In this work, we propose an OMI system that could provide the room-scale locomotion input and the body motion input simultaneously. Our system only needs 2m2 area and the time delay less than 20ms. The experiments demonstrate the effectiveness of our OMI system.

Partial shading and hot spots may cause power loss and sometimes irreversible damage of photovoltaic (PV) modules. In order to evaluate the power generation of PV modules, it is necessary to calculate the area of shading and hot spots. The perspective distortion makes region closer to the camera have more pixels, which results in misestimate of shading area and hot spots. To solve this problem, a robust automatic method for removing projective distortion of photovoltaic modules from close shot images is proposed in this paper. Firstly, Images are converted to gray scale, and their edges are detected by Canny algorithm. Then, lines are detected by Hough transform, vanishing points are found by intersecting lines, and the line at infinity is specified by vanishing points. Next, the projective transformation matrix is decomposed into an affine transformation matrix and a simple projective transformation matrix, which has only two degrees of freedom. Affine rectification is derived from computing this simple projective transformation matrix, and finally, right angles are recovered by computing shear transformation matrix. The close shot visible and infrared images are collected for experiments. The results show that the proposed method performs better on rectification of PV modules from close shot images.

Early event detection is among the keys in the field of event detection due to its timeliness in widespread applications. The objective of early detection (ED) is to identify the specified event of the video sequence as early as possible before its ending. This paper introduces semi-supervised learning to ED, which is the first attempt to utilize the domain knowledge in the field of ED. In this setting, some domain knowledge in the form of pairwise constraints is available. Particularly, we treat the segments of complete events as must-link constraints. Furthermore, some segments do not overlap with the event are put together with the complete events as cannot-link constraints. Thus, a new algorithm termed semi-supervised ED (SemiED) is proposed, which could make better early detection for videos. The SemiED algorithm is a convex quadratic programming problem, which could be resolved efficiently. We also discuss the computational complexity of SemiED to evaluate its effectiveness. The superiority of the proposed method is validated on two video-based datasets.

Wind power prediction is important for smooth power generation from wind turbines. Due to the characteristics of volatility and indirectness of wind power, it is difficult to achieve high-precision prediction. In real world, the data collected from wind farms may be incomplete because of measurement error, malfunctioning sensors, misoperation, etc. Therefore, it is intractable to consider the wind power prediction problem under the missing data scenario. This paper focuses on this problem in a wind farm by combining Gaussian process regression and multiple imputation approaches. Expectation-maximization algorithm is used for estimating mixture components of the underlying data distribution as well as for dealing with missing data. New data sets are generated via multiple imputation strategy. Gaussian process regression models are constructed to perform predictions for each of these data sets. With ensemble approach, a final predictive model is developed. Experimental results demonstrate that, comparing with several existing approaches, the proposed method is effective for wind power prediction with missing data.

Graph-based methods have been widely adopted to detect salient objects in images. However, there are two limitations of these methods. First, only one kind of query is employed for saliency propagation on the graph. Second, these methods only represent pairwise relations between vertices and thus give an incomplete representation of the relationships between image regions. In this paper, we propose a foreground- and background-queries-based hypergraph optimization framework for salient region detection. In this framework, both foreground queries and background queries are explicitly exploited to uniformly highlight the salient foreground and suppress the non-salient background. Furthermore, to include both the pairwise and the higher-order relations among two or more vertices, a probabilistic hypergraph is constructed based on local spatial correlation, global spatial correlation, and color correlation to represent the relations among image regions from different views. Extensive experimental results demonstrate the effectiveness of the proposed framework.

pointer meters are widely used in gas gathering stations, and manual meter reading is time-consuming and laborious. This paper proposes an automatic reading method of pointer meters in gas gathering stations based on deep learning and image processing algorithm. Firstly, the input image is detected by YOLO neural network algorithm to quickly locate the position of the pointer meter in the image; then, the meter image is segmented from the background image, and after filtering and binarization preprocessing, the meter border is preliminarily detected by Hough transform. After further removing the background interference, the pointer area and scale value area are segmented by connected domain detection to detect. After measuring the center of mass of the scale value area, the center of the dial is corrected again by RANSAC fitting circle, and the pointer angle is identified by index table thinning algorithm, and finally the meter indication is calculated. The results show that this method can quickly locate the pointer meters in the gas gathering station under complex background and reduce the light interference.

The cloud cover information in ground-based cloud (GBC) images is important to direct normal irradiance (DNI) prediction. In order to obtain better performance, the DNI forecasting model needs to incorporate the features from GBC images and numerical time series. In this paper, to resolve the issues of feature fusion with various types, three novel fuzzy inference systems (FIS) models are proposed for intra-hour DNI prediction. The features of the GBC image and the numerical time series are firstly fuzzified by clustering and grid partition, respectively. Then, the hybrid fuzzification and optimization method lead to the presented two hierarchical fuzzy inference systems (HFIS) models and an adaptive neuro-fuzzy inference system (ANFIS) model for DNI prediction. The performance of the proposed models is validated with the data from the National Renewable Energy Laboratory (NREL) from January 1, 2018 to December 31, 2018. Experiments demonstrate that the proposed models outperform the reference model, and the best model is capable of achieving 19.66% improvement over the reference model for 10-minute ahead DNI prediction.

Catastrophic faults in photovoltaic (PV) arrays, if not detected timely, will significantly reduce the output power and even cause fire risks. Thus, fast and accurate fault detection and diagnosis techniques are of importance to enhance the efficiency, reliability, and safety of PV systems. However, conventional methods fail to detect faults at low mismatch levels and low irradiance levels. In this paper, a sensorless detection technique is thus proposed based on the instantaneous current reduction between two maximum power point tracking (MPPT) sampling instants. Simulations have been conducted to validate its availability to identify anomalies under varying environments, irrespective of the mismatch and irradiance levels. The results clearly demonstrate that the proposed method is less time-consuming but more accurate than the existing techniques. However, due to the presence of blocking diodes, the similarity between faults and certain partial shading (PS) conditions will make them more difficult to be classified. Hence, a unique point on the current–voltage (I-V) curve, namely inflection point, is exploited to accurately distinguish PS conditions, and further determine the mismatch level. More simulation tests under different conditions reveal that the proposed method is effective to quantitatively evaluate the faults. Accordingly, different voltages can be set to differentiate the currents in the faulty string and normal ones as much as possible for efficient fault location.

Wind speed presents a potential seasonal pattern revealed by the self-similarity in wavelet periodogram with various scales. The corresponding seasonal pattern will promote the improvement of the short-term wind speed forecasting accuracy. In this study, a novel method for short-term wind speed forecasting using wavelet transformation (WT) and AdaBoost technique is proposed to analyse the wind speeds distribution features and promote the model configuration. Power spectrum and seasonal pattern analysis using the WT are presented to investigate the wind speeds feature distribution based on the scalogram percentage of energy distribution in different seasons. This procedure contributes to perfecting the investigation of wind speed seasonal pattern characteristics over time and promotes the sample division by computing the statistics measurement based on the estimated frequencies interval. The model order estimation based on the information criteria is processed to reflect the systems dynamical sustainability between the current outputs and historical data. Finally, the experiments based on the real data from Yunnan wind farm are given to verify the effectiveness of the proposed approach.

Time series forecasting is one of the most important issues in numerous applications in real life. The objective of this study was to propose a hybrid neural network model based on wavelet transform (WT) and feature extraction for time series forecasting. The motivation of the proposed model, which is called PCA-WCCNN, is to establish a single simplified model with shorter training time and satisfactory forecasting performance. This model combines the principal component analysis (PCA) and WT with artificial neural networks (ANNs). Given a forecasting sequence, order of the original forecasting model is determined firstly. Secondly, the original time series is decomposed into approximation and detail components by employing WT technique. Then, instead of using all the components as inputs, feature inputs are extracted from all the sub-series obtained from the above step. Finally, based on the extracted features and all the sub-series, a famous neural network construction method called cascade-correlation algorithm is applied to train neural network model to learn the dynamics. As an illustration, the proposed model is compared with two classical models and two hybrid models, respectively. They are the traditional cascade-correlation neural network, back-propagation neural network, wavelet-based cascade-correlation network using all the wavelet components as inputs to establish one model (WCCNN) and wavelet-based cascade-correlation network with combination of each sub-series model (WCCNN multi-models). Results obtained from this study indicate that the proposed method improves the accuracy of ANN and can yield better efficiency than other four neural network models.

For years, researchers in neural network (NN) area have been carried out much productive research in improving the generalization ability of NNs. In this paper, a novel neural network design algorithm is presented for solving multi-class problems, structure decomposition based on Skeletonization (SDBSkeletonization), which is to simplify NNs further. The proposed method decomposes a complex multi-class problem into a set of two-class problems, each of which can be regarded as an individual problem. After learning all these individual problems in parallel with Skeletonization algorithm, we then integrate these results to final decision. In addition, Skeletonization solves the classification problem based on automatic feature extraction. This perspective gives a broader range of application of our method. Our experimental results on Waveform and Handwritten Digits database demonstrate that SDBSkeletonization improves the overall classification performance.

The alternating direction method has been used widely in the power systems field for solving the multiarea dispatch problem. However, experience with applications has shown that the convergence rate of the alternating direction method depends significantly on the selection of the penalty parameter of the linear consistency constraint. Typically, it is difficult to obtain the optimal penalty parameter in advance. In this paper, for the purpose of solving this problem, we propose centralized and distributed self-adaptive penalty parameter strategies that allow the value of the penalty parameter to increase or decrease based on the information from each iteration. Simulation results illustrate that the proposed centralized and distributed self-adaptive methods are superior to the traditional alternating direction method in terms of robustness and convergence rate.

© 2012 IEEE. Video-based facial expression recognition has received substantial attention over the past decade, while early expression detection (EED) is still a relatively new and challenging problem. The goal of EED is to identify an expression as quickly as possible after the expression starts and before it ends. This timely ability has many potential applications, ranging from human-computer interaction to security. The max-margin early event detector (MMED) is a well-known ranking model for early event detection. It can achieve competitive EED performance but suffers from several critical limitations: 1) MMED lacks flexibility in extracting useful information for segment comparison, which leads to poor performance in exploring the ranking relation between segment pairs; 2) the training process is slow due to the large number of constraints, and the memory requirement is also usually hard to satisfy; and 3) MMED is linear in nature, and hence may not be appropriate for data in a nonlinear feature space. To overcome these limitations, we propose an online multi-instance learning (MIL) framework for EED. In particular, the MIL technique is first introduced to generalize MMED, resulting in the proposed MIL-based EED (MIED), which is more general and flexible than MMED, since various instance construction and combination strategies can be adopted. To accelerate the training process, we reformulate MIED in the online setting and develop online multi-instance learning framework for EED (OMIED). To further exploit the nonlinear structure of the data distribution, we incorporate the kernel methods in OMIED, which results in the proposed online kernel multi-instance learning for early expression detection. Experiments on two popular and one challenging video-based expression data sets demonstrate both the efficiency and effectiveness of the proposed methods.

In this paper, we propose a new method for 3D hand scan registration. This method includes two stages: liner blending skinning(LBS) and non-rigid ICP. A template hand mesh model is used to register to the 3D hand scans, which have large pose and shape variations. To fit different types of 3D hand scans, the template model is driven by the skeletons follow the principle of LBS to eliminate pose variations, and then non-rigid ICP is used to deform the template so that shape variations can be eliminated. Finally, template and scan model completely overlap together, and the vertices of template and scan can be matched one by one. We believe that the ability to align a template to each scan is the basis for more in-depth research in future. For example, this work can be used as basis for training pose and shape parameters. With the help of these parameters a template can be used to model any 3D hand in real life just like SMPL[7] for human and SMAL[14] for animals.

A new behavioral model of nonlinear RF power amplifiers (PA) is presented to treat memory effects. The key issue here is considering the phenomenon that uplink quadrature modulation signals are influenced not only by the current uplink signals, but also by previous terms. The variation of AM/AM and AM/PM, and the asymmetries in lower and upper intermodulation terms are frequently observed in high-power PAs. To treat these phenomena, this paper proposes a model based on artificial neural network. The contribution made of this model is to solve the order determination issue (determining the order of the previous output and input signals that have influence on the current output signal). In addition, the recognized difficulty of long training process is overcome by using SDBCC algorithm, a novel neural network design method combining structure decomposition and the Cascade-Correlation neural network algorithm. The required maximum delay is established by examining the autocorrelation coefficient of the residual error. Ensemble system is finally used to improve the performance further. This proposed method is successfully validated in nonlinear modeling of the RF PAs from HuaWei Company, including 8000 samples.

The multilayer perceptrons (MLPs) have been widely used in many communication applications, however, the learning process of the multilayer perceptrons often becomes very slow, which is due to the existence of the singularities in the parameter space. As the singularities significantly affect the learning dynamics of MLPs, the standard gradient descent method is not Fisher efficient. In order to overcome this problem, natural gradient method and adaptive natural gradient method were proposed to accelerate the learning process. As is well known, step size in each iteration plays a key role in the performance of algorithms. In this paper, the modified adaptive natural gradient method is proposed where the step size in each iteration is adaptive modified. The aim of the proposed algorithm is to accelerate the convergence speed and increase the performance of MLPs. The simulation results verify the validity of the analytical results.

Data centers have been widely used in many pervasive computing applications. This paper shows the cost-aware virtual machine allocation problem for off-grid green data centers is an integer programming problem which is NP-hard. The paper presents a cost-aware virtual machine allocation algorithm which attempts to utilize renewable energy sources and minimize the energy cost of fossil fuel while maintaining the quality-of-service requirements of the tasks. The simulation results show that the proposed algorithm is sensitive to the changes in the price of fossil fuel, and is able to achieve scalable performance.

In recent years, many graph-based methods have been introduced to detect saliency. These methods represent image regions and their similarity as vertices and edges in a graph. However, since they only represent pairwise relations between vertices, they give an incomplete representation of the relationships between regions. In this work, we propose a hypergraph based optimization framework for salient object detection to include not only the pairwise but also the higher-order relations among two or more vertices. In this framework, besides the relations among vertices, both the foreground and the background queries are explicitly exploited to uniformly highlight the salient objects and suppress the background. Furthermore, a probabilistic hypergraph is constructed based on local spatial correlation, global spatial correlation, and color correlation to represent the relations among vertices from different views. Extensive experiments demonstrate the effectiveness of the proposed method.

The accurate estimation of direct normal irradiance (DNI) under clear sky conditions plays an important role in the concentrated solar thermal plant. A hybrid model with adjustable inputs is proposed to calculate the clear-sky DNI, including a base clear-sky model and an error-correction model. The base clear-sky model is able to estimate the clear-sky DNI at any place with only the local date and location information, and the error-correction model serves as a supplementary to improve the calculating accuracy with available meteorological data. The error-correction model effectively integrates a linear part and a nonlinear part, and its inputs are adjustable according to the available meteorological observations. Several experiments have been conducted to evaluate the performance of the proposed model with data from three observation stations provided by the National Renewable Energy Laboratory open database. The results show that the hybrid model is able to provide great improvement over the base clear-sky model with 28%–70% on normalized root mean square error, and it also performs better than those using a linear or nonlinear error correction model. It is concluded that the performance of the hybrid model is comparable with other published methods in calculating the clear-sky DNI with concrete statistics.

Wind speed prediction is the key to wind power prediction, which is very important to guarantee the security and stability of the power system. Due to dramatic changes in wind speed, it needs high-frequency sampling to describe the wind. A large number of samples are generated and affect modeling time and accuracy. Therefore, two novel active learning methods with sample selection are proposed for short-term wind speed prediction. The main objective of active learning is to minimize the number of training samples and ensure the prediction accuracy. In order to verify the validity of the proposed methods, the results of support vector regression (SVR) and artificial neural network (ANN) models with different training sets are compared. The experimental data are from a wind farm in Jiangsu Province. The simulation results show that the two novel active learning methods can effectively select typical samples. While reducing the number of training samples, the prediction performance remains almost the same or slightly improved.

Summary In this paper, a novel with respect to economic dispatch of large-scale multi-area power systems is formulated and studied based on decomposition and coordination algorithm. More precisely, auxiliary problem principle (APP) is applied to overcome the defects of augmented Lagrangian relaxation technology with nonlinear inseparable quadratic terms. The main purpose of this paper is to investigate the intrinsic relationship between APP control parameters and convergence speed, and develop a general criterion that ensures that a suitable control parameter set referring to the implemented APP can be effectively estimated. Theoretical and experimental results of this paper are consistent with each other based on the verification of proposed numerical results. Copyright © 2013 John Wiley & Sons, Ltd.

Video-based facial expression recognition (FER) has recently received increased attention as a result of its widespread application. Using only one type of feature to describe facial expression in video sequences is often inadequate, because the information available is very complex. With the emergence of different features to represent different properties of facial expressions in videos, an appropriate combination of these features becomes an important, yet challenging, problem. Considering that the dimensionality of these features is usually high, we thus introduce multiview dimension reduction (MVDR) into video-based FER. In MVDR, it is critical to explore the relationships between and within different feature views. To achieve this goal, we propose a novel framework of MVDR by enforcing joint structured sparsity at both inter- and intraview levels. In this way, correlations on and between the feature spaces of different views tend to be well-exploited. In addition, a transformation matrix is learned for each view to discover the patterns contained in the original features, so that the different views are comparable in finding a common representation. The model can be not only performed in an unsupervised manner, but also easily extended to a semisupervised setting by incorporating some domain knowledge. An alternating algorithm is developed for problem optimization, and each subproblem can be efficiently solved. Experiments on two challenging video-based FER datasets demonstrate the effectiveness of the proposed framework.

Feature selection is a significant preprocessing procedure for many high-dimensional forecasting problems. In multivariate time series forecasting, the purpose of feature selection is to select a relevant feature subset from the original time series. One of the most common strategies for feature selection is mutual information (MI) criterion. K-nearest neighbors (k-NN) is a powerful means which has been widely used to estimate the MI between two high-dimensional variables directly from data set. In order to improve the effect of feature selection, a novel procedure is proposed in this paper to use MI and k-NN to perform feature selection in multivariate time series forecasting. This feature selection procedure not only reduces the irrelevance between the inputs and outputs variables, but also removes the redundant input variables. The procedure of this method is elaborated on a synthetic data set as well as on a real-world example.