Search

With the rapid growth of multimedia data (e.g., image, audio, and video) on the Web, the learning-based hashing techniques, such as deep supervised hashing, have proven to be very efficient for large-scale multimedia search. The recent successes seen in the learning-based hashing methods are largely due to the success of the deep learning-based hashing methods. However, there are some limitations to the previous learning-based hashing methods (e.g., the learned hash codes containing repetitive and highly correlated information). In this paper, we propose a novel learning-based hashing method, named deep attention-guided hashing (DAgH). DAgH is implemented using two stream frameworks. The core idea is to use the guided hash codes which are generated by the hashing network of the first stream framework (called the first hashing network) to guide the training of the hashing network of the second stream framework (called the second hashing network). Specifically, in the first network, it leverages an attention network and hashing network to generate the attention-guided hash codes from the original images. The loss function we propose contains two components: the semantic loss and the attention loss. The attention loss is used to punish the attention network to obtain the salient region from pairs of images; in the second network, these attention-guided hash codes are used to guide the training of the second hashing network (i.e., these codes are treated as supervised labels to train the second network). By doing this, DAgH can make full use of the most critical information contained in images to guide the second hashing network in order to learn efficient hash codes in a true end-to-end fashion. Results from our experiments demonstrate that DAgH can generate high-quality hash codes and it outperforms the current state-of-the-art methods on three benchmark datasets: CIFAR-10, NUS-WIDE, and ImageNet.

Due to its fast retrieval and storage efficiency capabilities, hashing has been widely used in nearest neighbor retrieval tasks. By using deep learning-based techniques, hashing can outperform non-learning-based hashing technique in many applications. However, we argue that the current deep learning-based hashing methods ignore some critical problems (e.g., the learned hash codes are not discriminative due to the hashing methods being unable to discover rich semantic information and the training strategy having difficulty optimizing the discrete binary codes). In this paper, we propose a novel image hashing method, termed as asymmetric deep semantic quantization (ADSQ). The ADSQ is implemented using three stream frameworks, which consist of one LabelNet and two ImgNets. The LabelNet leverages the power of three fully-connected layers, which are used to capture rich semantic information between image pairs. For the two ImgNets, they each adopt the same convolutional neural network structure but with different weights (i.e., asymmetric convolutional neural networks). The two ImgNets are used to generate discriminative compact hash codes. Specifically, the function of the LabelNet is to capture rich semantic information that is used to guide the two ImgNets in minimizing the gap between the real-continuous features and the discrete binary codes. Furthermore, the ADSQ can utilize the most critical semantic information to guide the feature learning process and consider the consistency of the common semantic space and Hamming space. The experimental results on three benchmarks (i.e., CIFAR-10, NUS-WIDE, and ImageNet) demonstrate that the proposed ADSQ can outperform current state-of-the-art methods.

The walking state monitoring is indispensable during the robot-aided walking of people with lower limb dysfunctions. In this paper, the existence of human-robot coordination state is first statistically verified in the process of using a walking-aid cane-type robot during walking. Based on this coordination, a new walking state monitoring method is proposed by using the principal component analysis (PCA). The abnormal or emergency walking state is promptly detected if the new sample data are found to deviate from an off-line PCA model, which is generated from plentiful normal walking data of different subjects. Furthermore, a state diagnosis algorithm based on the contribution plot is also developed for the walking state recognition and diagnosis. In this way, typical abnormal states like the leg restrictions can be distinguished from the emergency states including falls and the stumbling. Moreover, the human-robot coordination analysis can be performed using less sensors built-in the robot without needing the posture information of full human body. The effectiveness of the proposed method is proven by experiments. Better recognition rate and real-time performance of the method are also verified by comparing with conventional center of pressure based monitoring method.

Deep convolutional neural networks (DCNNs) are currently popular in human activity recognition (HAR) applications. However, in the face of modern artificial intelligence sensor-based games, many research achievements cannot be practically applied on portable devices (i.e., smart phone, VR/AR). DCNNs are typically resource-intensive and too large to be deployed on portable devices, and thus, this limits the practical application of complex activity detection. In addition, since portable devices do not possess high-performance graphic processing units, there is hardly any improvement in Action Game (ACT) experience. Besides, in order to deal with multi-sensor collaboration, all previous HAR models typically treated the representations from different sensor signal sources equally. However, distinct types of activities should adopt different fusion strategies. In this paper, a novel scheme is proposed. This scheme is used to train 2-bit CNNs with weights and activations constrained to {-0.5, 0, 0.5}. It takes into account the correlation between different sensor signal sources and the activity types. This model, which we refer to as DFTerNet, aims at producing a more reliable inference and better trade-offs for practical applications. It is known that quantization of weights and activations can substantially reduce memory size and use more efficient bitwise operations to replace floating or matrix operations to achieve much faster calculation and lower power consumption. Our basic idea is to exploit quantization of weights and activations directly in pre-trained filter banks and adopt dynamic fusion strategies for different activity types. Experiments demonstrate that by using a dynamic fusion strategy, it is possible to exceed the baseline model performance by up to ~5% on activity recognition data sets, such as the OPPORTUNITY and PAMAP2 data sets. Using the quantization method proposed, we were able to achieve performances closer to that of the full-precision counterpart. These results were also verified using the UniMiB-SHAR data set. In addition, the proposed method can achieve ~9x acceleration on CPUs and ~11x memory saving.

A synchronous condenser (SC) can provide the dynamic reactive power for the ultra-high-voltage direct current converter station during the system voltage drop caused by the single-phase short-circuit fault, which is one of the prevalent faults. However, the single-phase short-circuit fault can also result in the increase of the rotor loss and temperature and thus restrict the transient operating ability of the SC. A key factor that affects the rotor loss and temperature rise is the material of the rotor slot wedges. For a 300-MVar SC with the rotor slot wedges made of the aluminum, beryllium bronze, and stainless steel, the rotor loss and temperature distribution under the single-phase short-circuit are calculated by coupling the electromagnetic and temperature field models of the SC with the models of the power grid. The variations of the rotor losses along with the rotor slot wedge conductivity are revealed. The rotor regions, where the highest temperature rise appears, are obtained. According to the maximum permissible temperature of the rotor material, the transient operating ability of the SC withstanding the single-phase short-circuit fault is obtained. This article can provide the theoretic basis for improving the transient operating ability of the SC. [ABSTRACT FROM AUTHOR]

A short period of steady-state asynchronous operation after loss of excitation can improve the reliability of turbogenerator. The rotor damping structure of the turbogenerator can provide asynchronous torque and, thus, play an important role in improving the asynchronous operating ability. The rotor slot wedges as a part of rotor damping structure often made of aluminum, beryllium bronze, or stainless steel. The different materials of the rotor slot wedges can affect not only the asynchronous operating ability of the turbogenerator but also the rotor losses. The selection of material for the rotor slot wedges is therefore a key factor that should be carefully considered in the design of turbogenerator. In this article, the effect of the rotor slot wedges made of different materials on the asynchronous operating ability, rotor loss densities, and loss distributions of the turbogenerator after loss of excitation is investigated by time-stepping finite element model. The variations of the asynchronous operating ability and the rotor losses are revealed along with the rotor slot wedge conductivity. The rotor regions where the losses are significantly affected by the rotor slot wedge material are obtained. The results provide theoretic basis for improvement of the asynchronous operating ability and reduction of the rotor loss of turbogenerator. [ABSTRACT FROM AUTHOR]

A dual-excited synchronous generator (DESG) with two field windings on the rotor has better dynamic characteristics and high stability. However, the realization of the function needs a reasonable excitation control strategy. In this article, a power tracking excitation control (PTEC) is proposed to improve the operating performances of the DESG. The dynamic processes of a 300-MW DESG with the PTEC are calculated after the disturbance of the torque and reactive power by the time-stepping finite element model, and the results are superior to those adopting the traditional dual-channel excitation control. To further improve the dynamic characteristic of the DESG, the sensitivities of the control parameters of the PTEC to the dynamic characteristic indices are studied. The fuzzy control rules are built based on the sensitivity analysis results and the control parameters are optimized by the fuzzy control method. It revealed the variation laws of the optimized control parameters along with the active and reactive power deviations and their variable rates. The results show that the DESG adopting the optimized control parameters has better dynamic characteristics. The study can provide the theoretic basis for enhancing the system damping effect and suppressing the system oscillation. [ABSTRACT FROM AUTHOR]

A magnetic shield is widely used in the end region of a synchronous condenser (SC) to decrease the flux densities and losses of the end structural components. In order to investigate the influence of different magnetic shield thicknesses on the distributions of the flux density and loss in the end region, a 300-Mvar SC is selected as the investigation reference in this article. A three-dimensional (3-D) mathematical model used for the calculation of electromagnetic field and loss in the end region of the SC is established. The flux densities and losses of the end structural components are calculated and compared under the different thicknesses of the magnetic shield. From the detailed performance evaluations by the 3-D finite-element analysis, the variation of the flux densities and loss densities in the end region of the SC along with the thickness of the magnetic shield is revealed. The loss distributions of the main end structural components are studied and the end zones where the loss is most significantly affected by the magnetic shield thickness are obtained. The results provide the theoretical basis for the improvement of the loss distribution in the end region of the SC. [ABSTRACT FROM AUTHOR]

The attention mechanism of the Transformer has the advantage of extracting feature correlation in the long-sequence data and visualizing the model. As time-series data, the spatial and temporal dependencies of the EEG signals between the time points and the different channels contain important information for accurate classification. So far, Transformer-based approaches have not been widely explored in motor-imagery EEG classification and visualization, especially lacking general models based on cross-individual validation. Taking advantage of the Transformer model and the spatial-temporal characteristics of the EEG signals, we designed Transformer-based models for classifications of motor imagery EEG based on the PhysioNet dataset. With 3s EEG data, our models obtained the best classification accuracy of 83.31%, 74.44%, and 64.22% on two-, three-, and four-class motor-imagery tasks in cross-individual validation, which outperformed other state-of-the-art models by 0.88%, 2.11%, and 1.06%. The inclusion of the positional embedding modules in the Transformer could improve the EEG classification performance. Furthermore, the visualization results of attention weights provided insights into the working mechanism of the Transformer-based networks during motor imagery tasks. The topography of the attention weights revealed a pattern of event-related desynchronization (ERD) which was consistent with the results from the spectral analysis of Mu and beta rhythm over the sensorimotor areas. Together, our deep learning methods not only provide novel and powerful tools for classifying and understanding EEG data but also have broad applications for brain-computer interface (BCI) systems. [ABSTRACT FROM AUTHOR]