Search

In the field of super-resolution reconstruction, generative adversarial networks are able to generate textures that are more in line with the perception of the human eyes, but low-resolution images often encounter information loss and edge blurring problems in the process of reconstruction. In order to solve this problem, this article proposed an image super-resolution reconstruction model based on an enhanced attention mechanism and gradient correlation loss, which can better focus on important details in low-resolution images, thus improving the quality of reconstructed images. Firstly, an enhanced attention mechanism is proposed and incorporated into the generator model as a way to reduce the amount of information loss during image feature extraction and retain more image details. Furthermore, this paper proposed a gradient correlation loss function to maximize the correlation between the gradient of the generated image and the gradient of the original image. Thus, the generated image is more realistic and maintains a consistent edge structure. Finally, the experimental results on the standard dataset show that compared with other representative algorithms, the proposed algorithm has achieved some improvement in PSNR, SSIM, and LPIPS, which can verify the effectiveness of the algorithm.

Electroencephalography (EEG) plays a key role in the clinical evaluation of epilepsy and provides strong support for treatment decisions. However, analyzing and decoding EEG recordings is a burdensome task for neurologists and clinical experts. Existing automated detection techniques make considerable efforts in feature engineering, but often fall short when it comes to representing complex patterns in EEG signals. Deep learning methods allow for higher-order representations and intricate pattern learning, experiencing explosive success in the performance of automated EEG diagnostics. In this paper, we propose a novel Double Discrete Variational AutoEncoder (D2-VAE) network to learn the latent representations in EEG signals and perform deep discretization. Specifically, the method builds a learnable codebook based on the Vector Quantized Variational AutoEncoder (VQ-VAE) to obtain a generic representation of the EEG signal. The discretization of the local patterns of the signal is obtained by codebook queries, whereas the discretization of the global information of the signal is characterized by building a histogram of the quantization patterns. Such local-global signal portrayal is more attuned to the single-mode repetition and multi-mode mixing that characterizes epileptic signals. Multiple epilepsy diagnostic tasks and multiple evaluation metrics are used to validate the effectiveness of the proposed method for the classification of complex EEG signals. The experimental results show that D2-VAE possesses a low-dimensional yet powerful quantitative representation, with a significant improvement in performance over existing methods.

Legal case retrieval, which aims at retrieving similar legal cases based on the given query cases, plays an important role in intelligent legal systems. Existing methods including text-based and network-based methods have achieved notable advancements. However, text-based methods tend to overlook the significance of legal statutes and network-based information. Network-based methods ignore the rich text-based information in legal documents and are not applicable when legal citation networks are sparse. Moreover, most of the existing methods cannot be applied to inductive situations where new query cases continue to emerge. To overcome these issues, we propose a heterogeneous graph based on legal documents and legal statute hierarchy called LeDSGra, and utilize a heterogeneous graph representation learning method to combine both text-based and network-based information. Additionally, we design a module to identify relevant legal articles for query cases to connect the query cases to the existing citation network whereby the LeDSGra can be applied in inductive situations. Extensive experiments have been conducted on two real-world legal datasets which are LeCaRD and ELAM, and the experimental results demonstrate that our proposed method significantly improve the performance on the task of legal case retrieval. Moreover, our model outperforms other strong baseline models on both two datasets.

This paper investigates the interaction structure and decision-making process of swarm robots. It analyses the impact of communication links on the robots’ gains when they interact in local meta-structures and resolves the decision dynamic mechanisms of swarm robot systems. The collaboration efficiency of swarm robots is highly dependent on their local interactions. Complex networks are leveraged to portray the interaction structure between individuals, while classical games are employed to depict the decision-making paradigm of individuals. The study focuses on two typical local network meta-structures of distributed swarm robots: the cross-shaped grid world and the nine-lattice grid world. Within these meta-structures, the robots are categorized as resource-providing or resource-consuming. The information interaction and decision-making process between the central robot and its neighboring robots are mapped into resource-allocating group-interacting network games. The payoff of the central robot within the meta-structure is determined by the strategies of its neighbors and its neighbors’ neighbors, establishing the interaction strategies between robots, namely the yield game strategy and the price game strategy. A game model is constructed for meta-structures consisting of various classes and numbers of robots, which is subsequently solved. The results demonstrate that robots in network structures with simpler configurations and fewer communication links achieve higher gains in the group interaction network game. This economic perspective supports the notion that simpler network structures are more efficient and preferable to select.

In recent years, there has been a strong focus on the stability and control of stochastic networks, particularly single-layer networks. However, with the emergence of multi-layer networks, the research in this area has expanded significantly. Multi-layer networks have gained much attention due to their relevance in network science. In this paper, we address the synchronization of stochastic multi-layer networks (SMLNs) using finite-time (FnT) and fixed-time (FxT) methods with pinning control. Firstly, we provide the conditions for achieving FxT synchronization in SMLNs. Secondly, we present the conditions for FnT synchronization in SMLNs. Additionally, two kinds of finite-time convergent values of the system are given, which can be any given time. Furthermore, we discuss the relationship between the coupling matrix and the network layers, particularly in the context of minimum convergence time. This analysis has practical implications for real-world applications. Finally, we present numerical simulations to demonstrate the effectiveness of our proposed method.

The nozzle, converting pressure into kinetic energy, is the key component of fire extinguishing equipment, and its structure seriously affects the jet performance of fire monitors. In this study, the three-phase jet fire monitor (TPJFM) nozzle is taken as the research object, and the primary structural dimensions of the nozzle’s internal flow channel are considered as design parameters. A multi-objective optimization approach is employed to minimize the average turbulent kinetic energy at the outlet of the nozzle internal flow channel and the pressure drop between the inlet and outlet of the nozzle internal flow channel by integrating the Kriging surrogate model and NSGA-II. The research results indicate the pressure drop between the inlet and outlet of the nozzle internal flow channel decreases by 3.48%, and the average turbulent kinetic energy at the outlet of the nozzle internal flow channel reduces by 29.38% after optimization. The flow velocity distribution at the outlet of the internal flow channel of the optimized nozzle is more uniform, with an increase in jet range by 6.25%. Reasonable optimization of nozzle structure can significantly improve jet performance, and this method can provide a reference for designing nozzle structures.

Leakage current is one of the important parameters reflecting the operation status of distribution network-type surge arresters. At this stage, the polymagnetic current sensor has the advantages of miniaturization and high accuracy for leakage current measurement, but the complexity of electromagnetic interference in the field easily introduces more noise interference signals, which limits the performance of the polymagnetic current sensor for field application. To this end, this paper proposes a leakage current measurement method for distribution network-type surge arresters based on EEMD-SVD and low-rank RBF neural network methods. Firstly, the measured leakage current is decomposed by Ensemble Empirical Mode Decomposition (EEMD) to decompose the modal components containing eigenfrequencies, and then Singular Value Decomposition (SVD) is used to extract the non-eigenfrequency signals. Finally, a low-rank Radical Basis Function (RBF) neural network is used to approximate the leakage current signal after de-interference and combined with the Gaussian window function to remove the white noise interference. The experimental results show that the signal-to-noise ratio of the polymagnetic current sensor is improved by about 20dB and the maximum average absolute error is only 2.62%, which can truly reflect the leakage current of the network-type lightning arrester.

With the excessive selection for body weight (BW) and breast muscle weight (BMW) in broiler chickens, the incidence of leg diseases has gradually increased, which may lead to severe mortality, decreased productivity, and growth restrictions. Traditional methods for detecting leg diseases heavily rely on the interpretation of X-ray images by professionals and scoring methods for chicken gait. However, X-ray images of broiler chicken legs suffer from low background contrast and small, blurry lesion areas, posing significant challenges for traditional target detection methods. This paper proposes an improved algorithm based on the latest YOLOv8 for detecting leg diseases in X-ray images of chicken legs. In the feature extraction phase, Partial Convolution (PConv) is introduced to the C2f module, effectively reducing computational complexity while more accurately extracting spatial features. By incorporating Channel Prior Convolutional Attention (CPCA) into the network backbone, dynamic allocation of attention weights in both channel and spatial dimensions is achieved, preventing the loss of feature details caused by convolution iterations and enhancing the representation capability of small object features. The feature fusion stage introduces a novel Gather-Distribute mechanism (GD), effectively improving the inter-layer information exchange. Additionally, a Partial Convolution-based Shared Weight Detection Head (SharedPConv head) is introduced in the network head, making the model more lightweight and effectively alleviating the overfitting issue. Experimental results demonstrate that the improved method achieves a 7.2% increase in average precision, with a speed of 66.8fps, meeting real-time requirements and performing the detection task more accurately.

In response to the challenge that traditional convolutional neural networks face in effectively capturing the posture and spatial relationships of plant disease lesions on leaves, leading to issues of low recognition accuracy and poor robustness, this paper proposes a plant leaf disease image classification method that integrates capsule networks and residual networks. Firstly, by optimizing and refining the traditional residual network (ResNet), the initial convolutional layer of ResNet is enhanced by replacing its kernel with a concatenation of $3\times 3$ small convolutional kernels, aiming to more effectively extract features of plant leaf lesions. Subsequently, a channel attention mechanism is introduced into the residual block to heighten the model’s focus on crucial features. Finally, the improved ResNet is effectively integrated with the capsule network (CapsNet). The initial pooling layer of ResNet is removed to reduce the loss of positional information. The output of the third residual module of ResNet is then connected with CapsNet, fully leveraging the strengths of both networks to enhance the model’s robustness. Train and test the model on the PlantVillage, AI Challenger 2018, and Tomato Leaf Disease datasets and conduct comparative experiments with other typical classification models. The proposed SE-SK-CapResNet model has demonstrated a remarkable ability to accurately recognize diseased leaves, achieving an accuracy rate of 98.58%, 95.08%, and 97.19%, respectively. Furthermore, this model has exhibited superior performance in image rotation transformations and classification compared to traditional network models. These experimental results suggest that the SE-SK-CapResNet model is a promising solution for the detection of diseased leaves in the field of agriculture.

Fault diagnosis holds important significance in mitigating financial losses and ensuring equipment safety. As a crucial aspect of industrial machinery, bearing fault diagnosis becomes imperative. Nevertheless, in reality, identifying faults becomes challenging due to the presence of diverse variations in abnormal data, such as different vibration rotational speeds. In this paper, a novel framework for bearing fault diagnosis is proposed, called DTM-bearing, built upon the diffusion transformation model (DTM). This approach can transfer signals across different vibration speeds into a standardized signal aligned with a speed template. The primary purpose of DTM-bearing is to eliminate speed variations and extract speed-invariant features. Consequently, bolster the robustness of bearing fault diagnosis across diverse vibration speed scenarios. To the best of our knowledge, the proposed method is the first to combine the concepts of diffusion model and transformation in the domain of bearing fault diagnosis. Various experiments are performed on some datasets with multiple different speeds, which shows proposal can effectively improve the performance of bearing diagnosis. The framework based on the diffusion transformation model is expected to eliminate additional variations and improve the effectiveness of bearing diagnosis in practical applications.

Structural magnetic resonance imaging (sMRI) can identify subtle brain changes due to its high contrast for soft tissues and high spatial resolution. It has been widely used in diagnosing neurological brain diseases, such as Alzheimer’s disease (AD). However, the size of 3D high-resolution data poses a significant challenge for data analysis and processing. Since only a few areas of the brain show structural changes highly associated with AD, the patch-based methods dividing the whole data into several regular patches have shown promising for more efficient image analysis. The major challenges of the patch-based methods include identifying the discriminative patches, combining features from the discrete discriminative patches, and designing appropriate classifiers. This work proposes a novel efficient patch-based deep learning network (sMRI-PatchNet) with explainable patch localisation and selection for AD diagnosis. Specifically, it consists of two primary components: 1) A fast and efficient explainable patch selection method for determining the most discriminative patches; and 2) A novel patch-based network for extracting deep features and AD classification with position embeddings to retain position information, capable of capturing the global and local information of inter- and intra-patches. This method has been applied for the AD classification and the prediction of the transitional state moderate cognitive impairment (MCI) conversion with real datasets. The experimental evaluation shows that the proposed method can identify discriminative pathological locations effectively with a significant reduction on patch numbers used, providing better performance in terms of accuracy, computing performance, and generalizability, in contrast to the state-of-the-art methods.

The effective spoofing detection can prevent GNSS receiver from providing erroneous positioning, velocity and timing information. Most of the available multi-antennas carrier phase double difference-based spoofing detection methods, regardless of the aiding of inertial navigation system (INS), exploit the divergence in carrier phase double differences derived from the direction of arrival (DOA) characteristics of the authentic signal and spoofing signal. However, the traditional methods mainly implemented by dual-antenna GNSS receivers has spatial ambiguity and lacks global detection performance analysis. Furthermore, an INS-aided multi-antenna spoofing detection method is proposed without spatial ambiguity comparing the predicted signal DOA by INS and the vectoral carrier phase double differences measured by multi antennas. To analyze the influence of the antenna array baseline length and signal DOA on the detection performance, a comprehensive numerical simulation was conducted. The dead zone of the proposed detection method is revealed to assess the global spoofing detection performance. The effectiveness of the proposed method is verified under the GNSS positioning scenarios. In the worst case, spoofing signals can be effectively detected in 95.2% of the airspace under one time the wavelength antenna baseline configuration and in 99.3% of the airspace under five times the wavelength antenna baseline configuration, respectively.

A novel wideband quasi-Chebyshev bandpass filter (BPF) with steep skirt selectivity and extended upper stopband is presented. It is comprised of two microstrip-to-microstrip vertical transitions with nonuniform slotline sections, which cascaded directly by a nonuniform microstrip connecting line. In terms of odd-/even-mode analysis, the in-band six poles which make up the wideband bandwidth can be explicitly derived. Moreover, the Chebyshev equal-ripple filtering responses with prescribed specifications can be conveniently achieved in desired passband by adopting synthesis theory. In order to suppress the second and third spurious harmonics, a modified circuit model based on synthesis method is further proposed. Two additional transmission zeros (TZs) TZ3, TZ4 can be generated at $3f_{0}$ and $5f_{0}$ respectively by adjusting impedance ratio of slotline stubs properly under the condition of $\theta _{\mathrm {s1}} = \theta /4$ and $\theta _{\mathrm {s2}} = \theta /2$ , which widen the upper stopband bandwidth greatly and enhance harmonic rejection level effectively. Meanwhile, combing with the inherent periodic finite-frequency TZ5, TZ6 and TZ7, the upper stopband could be extended to $6.22f_{0}$ with rejection level better than 14.6 dB. Besides, a pair of finite-frequency TZ1 and TZ2 caused by source-load coupling occur at both sides of passband to sharpen the cutoff slope of skirts remarkably. For verification, a prototype is fabricated and measured, the experimental results agree well with the theoretical predictions, which verify the validity of proposed method.

It is well known that complex networks are inevitably affected by random disturbances. In the study of complex network stability, finite-time control has always been a research hot spot. This paper discuss finite-time synchronization (FTSY) of stochastic multilayer networks (MLNs). First, novel finite-time stability (FTST) theorems of deterministic complex dynamical system are given, and the expression of finite stability time is also estimated. Second, sufficient criteria are obtained for FTSY of stochastic MLNs. Compared with some the existing research results of FTSY in stochastic MLNs, FTSY of stochastic MLNs is realized by using FTST of deterministic complex dynamical systems, and the relationship between control intensity and network layers is also discussed under the condition of minimum convergence time, which will be meaningful to help how to select appropriate in actual complex system control. Lastly, the availability of the method is checked by numerical simulation base on the theory proposed in this paper.

In this study, with the help of a disturbance observer, an adaptive fractional-order sliding mode control scheme is designed to achieve trajectory tracking control of indeterminate manipulators. Firstly, the fractional-order non-singular fast terminal sliding mode (FO-NFTSM) surface is presented to increase the convergence velocity of the controller. Secondly, the adaptive reaching law is designed on the strength of the super-twisting algorithm to ensure the control performance of the approaching stage. Meanwhile, aiming at the compound disturbance existing in the manipulator system, an adaptive fixed-time sliding mode disturbance observer (AFSMDO) is introduced to settle the problem, which can estimate the lumped disturbance in real-time. Moreover, compensate for the system and raise the precision of the controller. The Lyapunov method is applied to certify the stability of the control system. Simulation experiments verify the superior property of the controller designed in this article.

Recently, Deep Neural Networks (DNNs) have recorded significant success in handling medical and other complex classification tasks. However, as the sizes of DNN models and the available datasets increase, the training process becomes more complex and computationally intensive, usually taking longer to complete. In this work, we have proposed a generic full end-to-end hybrid parallelization approach combining model and data parallelism for efficiently distributed and scalable training of DNN models. We have also proposed a Genetic Algorithm Based Heuristic Resources Allocation (GABRA) mechanism for optimal distribution of partitions on the available GPUs for computing performance optimization. We have applied our proposed approach to a real use case based on 3D Residual Attention Deep Neural Network (3D-ResAttNet) for efficient Alzheimer Disease (AD) diagnosis on multiple GPUs and compared with the existing state-of-the-art parallel methods. The experimental evaluation shows that our proposed approach is 20% averagely better than existing parallel methods in terms of training time and achieves almost linear speedup with little or no differences in accuracy performance when compared with the existing non-parallel DNN models.

Massive music data and diverse listening behaviors have caused great difficulties for existing methods in user-personalized recommendation scenarios. Most previous music recommendation models extract features from temporal relationships among sequential listening records and ignore the utilization of additional information, such as music’s singer and album. Especially, a piece of music is commonly created by a specific musician and belongs to a particular album. Singer and album information, regarded as music metadata, can be utilized as important auxiliary information among different music pieces and may considerably influence the user’s choices of music. In this paper, we focus on the music sequential recommendation task with the consideration of the additional information and propose a novel Graph-based Attentive Sequential model with Metadata (GASM), which incorporates metadata to enrich music representations and effectively mine the user’s listening behavior patterns. Specifically, we first use a directed listening graph to model the relations between various kinds of nodes (user, music, singer, album) and then adopt the graph neural networks to learn their latent representation vectors. After that, we decompose the user’s preference for music into long-term, short-term and dynamic components with personalized attention networks. Finally, GASM integrates three types of preferences to predict the next (new) music in accordance with the user’s music taste. Extensive experiments have been conducted on three real-world datasets, and the results show that the proposed method GASM achieves better performance than baselines.

The era of big data provides the possibility of precision medicine. The most important idea we have for cancer is to divide and treat. Theoretically, each person’s cancer should be different, so it is very necessary to make personalized treatment plans for different cancer patients. Subtype analysis of cancer can be viewed as a clustering problem, while ensemble clustering techniques are widely followed for their ability to combine multiple basic clusters into potentially better and more robust clusters. However, the reliability of the present ensemble clustering methods in cancer subtype analysis still needs to be improved. Therefore, we propose a double weighted ensemble clustering method (DWEC), which first derives the similarity matrix of each base cluster based on the local weighting method, and this process can be regarded as the first weighting based on clusters. Subsequently, the objective of finding the final partitions is regarded as an optimization problem, and the similarity matrix corresponding to each base cluster is weighted twice by the block coordinate descent algorithm to solve the optimal partitions result. The best experimental results were obtained in both labeled datasets and unlabeled cancer gene datasets, validating the superiority of the method. For cancer subtype analysis, although our proposed method did not show statistically significant differences in survival distributions of several subtypes in the subtype analysis of glioblastoma multiforme. However, it performed best in the results of the temporal test for all other four cancer gene data, and therefore, we conclude that our method is more effective for cancer subtype analysis compared with other methods.

Additive manufacturing (AM) of high-strength metals, which is typically based on laser powder bed fusion (LPBF), can introduce microscopic pores in the AM metal. Pulsed Infrared Thermography (PIT) offers several advantages for nondestructive imaging of subsurface defects in AM structures because the method is one-sided, non-contact and scalable to structures of arbitrary size. However, high-resolution PIT imaging results in the generation of a large volume of thermography data (~TB), which creates challenges for the storage and transmission of data. Compression of thermography data requires an approach that achieves high data compression ratio while preserving weak thermal features corresponding to microscopic material defects. We investigate thermography data compression using several unsupervised learning (UL) algorithms, which include Principal Component Analysis (PCA), Independent Component Analysis (ICA), Exploratory Factor Analysis (EFA), Sparse Dictionary Learning (SDL), and a novel lightweight Thermography Compressive Sparse Autoencoder (TCSA). Algorithms are benchmarked using PIT experimental data obtained from imaging of a stainless steel plate with calibrated porosity defects imprinted with AM process. For all algorithms, we obtain compression ratio >30 (highest compression of 46 is achieved with TCSA), and peak signal-to-noise ratio for reconstruction accuracy >73dB. Compared to existing methods, advantages of UL algorithms include achieving high compression ratio while preserving weak features to allow extraction of microscopic material defects from images. UL-based methods have general applicability because they are adaptable to compression of different data types, and allow for memory-efficient training and rapid on-line augmentation of the model.

Mobile edge computing (MEC) is considered a more effective technological solution for developing the Internet of Things (IoT) by providing cloud-like capabilities for mobile users. This article combines wireless powered communication (WPC) technology with an MEC network, where a base station (BS) can transfer wireless energy to edge users (EUs) and execute computation-intensive tasks through task offloading. Traditional numerical optimization methods are time-consuming approaches for solving this problem in time-varying wireless channels, and centralized deep reinforcement learning (DRL) is not stable in large-scale dynamic IoT networks. Therefore, we propose a federated DRL-based online task offloading and resource allocation (FDOR) algorithm. In this algorithm, DRL is executed in EUs, and federated learning (FL) uses the distributed architecture of MEC to aggregate and update the parameters. To further solve the problem of the non-IID data of mobile EUs, we devise an adaptive method that automatically adjusts the FDOR algorithm’s learning rate. Simulation results demonstrate that the proposed FDOR algorithm is superior to the traditional numerical optimization method and the existing DRL algorithm in four aspects: convergence speed, execution delay, overall calculation rate and stability in large-scale and dynamic IoT.

The point-of-load (POL) converters integrated with the input LC filter are essential to supply the ac load in the dc microgrid. The interaction between the LC filter and the POL converter will bring instability problems to the dc-link voltage. This paper proposes a series-connected virtual impedance regulator based on the detailed small-signal model of the POL converter that feeds a linear three-phase load. The proposed method can improve the stability by modifying only the phase of the input impedance of the POL converter within the interaction frequency range. Therefore, the impact of the regulator on the converter’s dynamic performance is minimized. The experiments show that the proposed stabilization methods exhibit good performance under different linear load scenarios.

This paper considers the theoretical limits of resource allocation overheads in vehicular ad hoc networks (VANETs), where the resource allocation procedure may consume extensive system resources. A permutation-based model is proposed to analyze the overhead limits where the unordered resource allocation patterns are considered as a permutation process. The minimum bits to represent the resource allocation information can be derived based on the permutation-based model. The bipartite graph is applied to describe the resource reuse in the respective domains for practical scenarios, and the dynamicity is taken into consideration. The proposed permutation-based model is applied to analyze the resource allocation overheads of two typical resource allocation schemes in VANETs: time-division multiple access (TDMA) and cluster-based algorithms. The proposed methodology provides new insights in a tradeoff analysis between overheads and system efficiency when analyzing the capacity of the VANETs.

Reinforcement learning has been widely applied for sequential decision making problems in various fields of the real world, including recommendation, e-learning, etc. The features of multi-policy, latent mixture environments and offline learning implied by many real applications bring a new challenge for reinforcement learning. To this challenge, the paper proposes a reinforcement learning approach called offline multi-policy gradient for latent mixture environments. The proposed method uses an objective of expected return of trajectory with respect to the joint distribution of trajectory and model, and adopts a multi-policy searching algorithm to find the optimal policies based on expectation maximization. We also prove that the off-policy technique of importance sampling and advantage function can be used by offline multi-policy learning with fixed historical trajectories. The effectiveness of our approach is demonstrated by the experiments on both synthetic and real datasets.

To tackle the inherent unknown deformation (e.g., translation, rotation and scaling) of the inverse synthetic aperture radar (ISAR) images, a deep polar transformer-circular convolutional neural network, i.e., PT-CCNN, is proposed to achieve deformation robust ISAR image automatic target recognition (ATR) in this article. Inspired by human visual system and canonical coordinate of Lie-groups, we adopt a polar transformer module to transform the deformation ISAR images to the log-polar representations, before which a conventional convolutional neural network (CNN) is utilized to predict the origin of log-polar transformation. The above techniques make the proposed network invariant to translation, and equivariant to rotation and scaling. On this basis, for the log-polar representations with wrap-around structure, a circular convolutional neural network (CCNN) is further applied to extract more effective and highly discriminative features and improve recognition accuracy. The proposed network is end-to-end trainable with a classification loss, and could extract deformation-robust and essential features automatically. For multiple practical ISAR image datasets of six satellites, the performance testing and comparison experiments demonstrate that the techniques utilized in PT-CCNN are effective, and our proposed network achieve higher recognition accuracy than those previous common methods based on deep learning. For instance, our proposed PT-CCNN beats traditional CNN on rotation, scaling and practical deformation datasets by 24.5-49.3%, 9.0-40.8% and 22.3-26.7%. And it also outperforms the polar CNN without using the above techniques on rotation, scaling and practical deformation datasets by 9.2-53.7%, 5.2-54.6% and 9.0-49.9%. Additionally, the presented visualization results show the abilities and advantages of our method in terms of handling image deformation and extracting effective features.

Traditional optical fiber vibration signal (OFVS) recognition research focuses on signal endpoint detection and feature extraction. These two aspects directly determine the success of OFVS recognition. The traditional method relies on artificially designed features and has a strong pertinence to the classification target, resulting in poor stability and flexibility. In response to the above problems, this paper combines the traditional OFVS recognition ideas (time-frequency analysis and feature extraction) and the characteristics of deep learning automatic learning parameters to construct an end-to-end adaptive filtering convolutional neural network (AF-CNN), which can directly get the classification results through the iterative update of the network. In modeling the original signal, the following steps were taken to make the network interpretable. First, we use a one-dimensional (1-D) convolution on the original OFVS. The convolution kernel can adaptively treat the original signal perform filtering to obtain filtered signals of different frequency bands. Second, using a general convolutional neural network (CNN) to extract the filtered signal features. Finally, a multi-layer perceptron (MLP) is used for classification. This paper compares the AF-CNN network with three traditional pattern recognition methods and proves that the AF-CNN network's accuracy is better than traditional pattern recognition methods. The average accuracy can reach 96.7%, and it can effectively distinguish OFVS with weaker energy and similar waveforms.

Ground penetrating radar (GPR) is a well-known useful tool for subsurface exploration. GPR data can be recorded at a relatively high speed in a continuous way with hyperbolas being artifacts and evidence of disturbances in the soil. Automatic and accurate detection and interpretation of hyperbolas in GPR remains an open challenge. Recently deep learning techniques have achieved remarkable success in image recognition tasks and this has potential for interpretation of GPR data. However, training reliable deep learning models requires massive labeled data, which is challenging. To address the challenges, this work proposes a Generative Adversarial Nets (GANs)-based deep learning framework, which generates new training data to address the scarcity of GPR data, automatically learns features and detects subsurface objects (via hyperbola) through an end-to-end solution. We have evaluated our proposed approach using real GPR B-scan images from rail infrastructure monitoring applications and compared this with the state-of-the-art methods for object detection (i.e. Faster-RCNN, Cascade R-CNN, SSD and YOLO V2). The proposed approach outperforms the existing methods with high accuracy of 97% being the mean Average Precision (mAP). Moreover, the proposed approach also demonstrates the good generalizability through cross-validation on independent datasets.

Service deployment often needs to guarantee QoS of service compositions for networks which compose services to complete users' application requests. Because evaluation methods for QoS of service compositions rely on fixed-process-based formal models which are not suitable for dynamic service composition processes in service deployment, current service deployment methods can only guarantee QoS of single service and are insufficient to guarantee QoS of service compositions. In this article, a service deployment method considering application reliability is proposed. First, application reliability is introduced to measure QoS of service compositions and an evaluation method based on application flow is presented. Then a service deployment optimization model with the goal of application reliability is established. To solve this model efficiently, an improved genetic algorithm by a new method generating initial population is adopted. Finally, 5G network case study verifies that the improved genetic algorithm converges quickly and our proposed method significantly outperforms the existing method in terms of application reliability by comparison.

Output LC filter is one of the most important parts for Buck converters. The existing optimization methods for LC filter fail to provide a fully optimized design. The difficulty in a holistic design approach lies in the trade-off relationships among different design targets. For example, smaller volume results in worse filtering capability and lower efficiency. To improve the overall performance of the output LC filter in Buck converter, a multi-objective design is proposed, taking the power loss, cut-off frequency and volume as design targets. This proposed holistic design approach utilizes Pareto-Frontier to achieve a compact LC filter with optimized efficiency and filtering capability. However, Pareto-Frontier generated by the previous multi-objective algorithms suffers from nonuniform or incomplete coverage, which seriously undermines design accuracy. Thus, the coevolving-AMOSA algorithm is proposed to provide a Pareto-Frontier with uniform and complete coverage. Via this proposed multi-objective design for the output LC filter in Buck converter with the coevolving-AMOSA algorithm, output LC filter can be flexibly designed to meet requirements in various applications while maintaining outstanding comprehensive performance. Optimal design cases for three specific application scenarios are presented as examples. Finally, the experimental results validate the effectiveness of the proposed multi-objective approach.

Images captured in low-light conditions often suffer from bad visibility, e.g., low contrast, lost details, and color distortion, and image enhancement methods can be used to improve the image quality. Previous methods have generally obtained a smooth illumination map to enhance the image but have ignored details, leading to inaccurate illumination estimations. To solve this problem, we propose a multi-scale attention retinex network (MARN) for low-light image enhancement, which learns an image-to-illumination mapping to obtain a detailed inverse illumination map inspired by retinex theory. In order to introduce more image priors, we introduce a novel illuminance-attention map to guide the model to characterize varying-lighting areas, which we combine with the low-light image as the model input. MARN consists of a multi-scale attention module and a feature fusion module; the former extracts multi-resolution features with attention-based feature aggregation, while the latter further merges the output features of the previous module with the input. To achieve better visibility, we formulate a novel loss function to synthetically measure the illumination, detail, and colorfulness of the image. Extensive experiments are performed on several benchmark datasets. The results demonstrate that our method outperforms other state-of-the-art methods according to both objective and subjective metrics.

This article provides a survey of recent results on exploiting parametric auto-regressive (AR) models for adaptive radar target detection. Specifically, three types of radar systems are considered, including phased-array radar with multiple co-located transmitters and receivers, distributed multi-input multi-output (MIMO) radar with widely and spatially separated transmitters and receivers, and passive radar which uses existing sources as illuminators of opportunity (IOs). These radar systems are of significant interest for a wide range of military and civilian applications. For each of the three types of radars, we discuss how AR processes can be employed to succinctly model the underlying signal correlation and efficiently estimate it from limited data, thus enabling effective target detection in complex non-homogeneous environments when training data is limited. We illustrate the performance of such parametric model assisted detectors relative to conventional non-parametric approaches by using computer simulated and experimental data.

Selective ground-fault protection is greatly valued for the safe and reliable operation of power systems. With the wide adoption of fault indicator in distribution network, the amount of available fault data increases dramatically. The in-depth investigation of fault recording data helps improve the accuracy of faulty line identification. To perform fault data analysis with higher efficiency, a single-phase-to-ground fault identification model based on the k-Nearest Neighbor (kNN) classification algorithm is proposed in the paper. In this model, the eigenvectors consist of wavelet energy ratio, wavelet coefficients variance and wavelet power obtained by the decomposition of transient components. Furthermore, through the theoretical analysis and experimental comparison of three parameter adjustment algorithms, Bayesian Optimization algorithm is selected to find the optimal parameters of fault identification model, and realize the adaptive adjustment of model parameters. Finally, the validity and feasibility of the model are verified by the experimental data, and the accuracy and efficiency of fault identification are improved by using Bayesian Optimization algorithm.

In view of the problem that the doctor cannot quickly obtain the three-dimensional information of the liver surface during the traditional open resection of liver tumor without radiation, a new three-dimensional reconstruction system of liver surface based on binocular structured light is proposed in this paper to obtain three-dimensional information of liver surface rapidly. The hardware construction and software module design of the three-dimensional reconstruction system of the liver surface are carried out. The multi-frequency heterodyne method is proposed to solve the problem of the lack of three-dimensional reconstruction of the liver surface. A GPU-CPU collaborative acceleration method is proposed to speed up the three-dimensional reconstruction of intraoperative liver surface. On this basis, a human respiratory platform is built to simulate the effect of human respiration on three-dimensional reconstruction of the liver surface in a real surgical environment. The accuracy and speed of the system are verified by the liver experiments: The errors of three-dimensional reconstruction of liver surface at expiratory peak and inspiratory peak are 0.78 ± 0.18mm and 0.92 ± 0.21 mm, respectively. The average time required for three-dimensional reconstruction of liver surface based on GPU-CPU collaborative calculation is 3.93s. The speed of three-dimensional reconstruction of liver surface based on GPU-CPU collaborative calculation is 52% faster than that of three-dimensional reconstruction of liver surface based on CPU. It can be seen that the system has high accuracy and fast speed, which can meet the basic needs of the doctor in surgery.

Automatic extraction of buildings from high-resolution remote sensing images becomes an important research. Since the convolutional neural network can perform pixel-level segmentation, this technology has been applied in this field. But the increase in resolution prone to blurry segmentation because the model needs more edge detail and multi-scale detail learning. To solve this problem, a method is proposed in this paper, which consists of three parts: (1) an improved model named Holistically-Nested Attention U-Net (HA U-Net) is designed, which integrates the attention mechanism and multi-scale nested modules to supervise prediction; (2) During model training, an improved weighted loss function is proposed to make the designed model more focused on learning boundary features; (3) watershed algorithm is exploited for image post-processing to optimize segmentation results. The designed HA U-Net performs well on WHU Building Dataset and Urban3d Challenge dataset, and achieves 9.31%, 2.17% better F1-score and 10.78%, 1.77% better IOU than the standard U-Net respectively. The experimental results indicate that the proposed method can well solve the building adhesion problem. The research can serve as updating geographic databases.

Higher requirements must be put forward about wireless signal transmission in harsh environments. Energy loss mitigation, channel estimation, noise interference reduction demands high quality of service in wireless communication. In the fifth generation wireless communication, massive multiple-input multiple-output (MIMO) networks have been used and play a significant role to fulfill the requirements. Cell-Free massive MIMO networks are recognized as possible solution in the future wireless communication. Spectral efficiency (SE) is a very important index in assessing massive MIMO networks. This article tries to optimize both SE and power control of cell-free massive MIMO networks. Both uplink and downlink transmission are included in the study. The optimization model contains both SE and power control of massive MIMO networks. To tackle the model, a novel ensemble method is developed inspired by ensemble learning. The ensemble method is built up on the neighborhood field optimization method. The goodness of the developed ensemble method is verified by comparing with genetic algorithm, gradient descent method and Newton method. Extensive simulations are performed to study the optimization of SE and power control. Different wireless sensors are taken into consideration to simulate different requirements of massive MIMO networks. The results demonstrate the effectiveness of the proposed method for applying in massive MIMO networks.

With the rapid development of the power grid, the high penetration of new energy sources and the diversity of loads have further aggravated the uncertainty of “source-load”, which has brought huge challenges to the peak shaving of the power grid. In order to ensure the reliable operation of the power system during the electricity peak, this paper combines the IoT technology to propose a flexible control strategy for controllable large industrial loads that considers the interrupt priority. Firstly, the perception and fusion framework of large industrial load information is constructed based on the IoT technology. After that, the improved TOPSIS method is adopted to establish the evaluation model of the adjustable potentials of large industrial loads and the load interruption priority is further divided. Finally, a three-stage rolling regulation model for controllable large industrial loads to participate in the deep peak shaving of the power grid is constructed to achieve the goal of bidirectional peak shaving on the power generation side and the demand side. The case uses an improved IEEE 30-node system for simulation. As the simulation results show, the method proposed in this paper can not only take the cost of peak load regulation into account, but also effectively achieve the goal of ‘peak shaving and valley filling’.

Random forest (RF) is an ensemble classifier method, all decision trees participate in voting, some low-quality decision trees will reduce the accuracy of random forest. To improve the accuracy of random forest, decision trees with larger degree of diversity and higher classification accuracy are selected for voting. In this paper, the RF based on Kappa measure and the improved binary artificial bee colony algorithm (IBABC) are proposed. Firstly, Kappa measure is used for pre-pruning, and the decision trees with larger degree of diversity are selected from the forest. Then, the crossover operator and leaping operator are applied in ABC, and the improved binary ABC is used for secondary pruning, and the decision trees with better performance are selected for voting. The proposed method (Kappa+IBABC) are tested on a quantity of UCI datasets. Computational results demonstrate that Kappa+IBABC improves the performance on most datasets with fewer decision trees. The Wilcoxon signed-rank test is used to verify the significant difference between the Kappa+IBABC method and other pruning methods. In addition, Chinese haze pollution is becoming more and more serious. This proposed method is used to predict haze weather and has achieved good results.

In recent years, the Sq-MRC (machine reading comprehension for separate questions) task, where the questioner poses a separate question each time, has experienced rapid development. However, we often encounter circumstances where the questioner poses a series of nonstop questions. The corresponding task, called Cq-MRC (machine reading comprehension for continuous questions), has been rarely investigated and is nontrivial due to (1) the incomplete expressions of subsequent questions, such as anaphora or subject ellipsis, and (2) the shortage of annotated samples for model training and performance evaluation. We explore this challenging task, in this article, and propose a completion-then-prediction approach. To tackle the first issue, we identify the key-entity in the first question and employ it to complete the subsequent ones, and hence decompose the task into a set of Sq-MRC tasks which are accomplished using a BERT-based model. In addition, to mitigate the shortage of annotated samples, we train a meta-learner on the Sq-MRC task which is related to Cq-MRC and has abundant labeled data, and then perform few-shot learning by fine-tuning the model on our target task, viz. Cq-MRC, using the limited number of samples available. The results of extensive experiments conducted on a multi-document Chinese machine reading comprehension dataset show the effectiveness and superiority of our method, achieving 2.3% to 3.7% absolute improvements in terms of major metrics.

Fog radio access networks (F-RANs) can effectively alleviate fronthaul loads and reduce content transmission delay by migrating cloud services to the network edge. This paper addresses a cooperative caching scenario in F-RAN, where each mobile user can acquire the requested contents from any one of its associated fog-computing-based access points (F-APs). However, caching disparate contents in different F-APs will lead to different content delivery delays, since mobile users suffer from diverse channel fadings and interferences when they download contents from different F-APs. Considering limited caching storage in each F-AP, diverse user preferences, unpredictable user mobility and time-varying channel states, an average transmission delay minimization problem is formulated. With the aid of dueling deep-Q-network framework, a delay-aware cache update policy is proposed for mobile users in F-RAN. The proposed cache update policy will decide to replace the stored contents in F-APs with the proper contents at each time slot. Compared with first in first out, least recently used and least frequently used caching policies, simulation experiments are performed to evaluate the performance of the proposed algorithm. Simulation results illustrate that the proposed caching policy yields better average hit ratio and lower average transmission delay than other traditional caching policies.

Achieving accurate navigation and localization is crucial for Autonomous Underwater Vehicle (AUV). Traditional navigation algorithms, such as Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF), require the system model and measurement model for state estimation to obtain the AUV position. However, this may introduce modeling errors and state estimation errors which will affect the final precision of AUV navigation system to a certain extent. To avoid these problems, in this paper, we proposed a deep framework - NavNet - by taking AUV navigation as a deep sequential learning problem. Firstly, the proposed NavNet can take raw sensor data at different frequencies as input, which benefits from the sequential learning capability of Recurrent Neural Network (RNN). Secondly, NavNet takes advantage of a simplified attention mechanism and Fully Connected (FC) layers to output AUV displacements per unit time, which accomplishes low-frequency AUV navigation by accumulation of it. More importantly, there is no need for the model building and state estimation with NavNet, which avoids the import of relevant errors. We compare the performance of NavNet to EKF and UKF using collected data by running Sailfish in the sea. Experimental results show that NavNet has an excellent performance in terms of both the navigation accuracy and fault tolerance. In addition, a reliable fusion strategy of NavNet and conventional method is applied to achieve high-frequency AUV navigation. The experimental results show that the proposed architecture can be a reliable supplement to limit the error growth of conventional algorithms.

To obtain chaotic systems with good properties to help chaos based image encryption, this paper first study the condition under which quadratic functions can be used as pseudo-random sequence generator. We give a necessary and sufficient condition for quadratic functions topologically conjugate with the most frequently used Logistic map and Tent map. Since topological conjugacy maintains chaotic characteristics, more quadratic functions can be used as pseudo-random sequence generators in the digital image encryption. Then, it analyzes a diffusion algorithm with a good effect but poor in resisting the chosen plaintext attack. It is also verified by simulation. The quadratic function chaotic system is applied in the improvement of the diffusion algorithm. Furthermore, the improved algorithm contains plaintext related and cyclic shift operations. Finally, the performance analysis shows that the improved scheme has good security and high efficiency.

Supply chain configuration (SCC) plays an important role in supply chain management. This paper focuses on a multi-objective SCC (MOSCC) problem for minimizing both the cost of goods sold and the lead time simultaneously. Some existing population-based methods use the evolution of a population to obtain the optimal Pareto set, but they are time-consuming. In this paper, an Efficient Local Search-based algorithm with rank (ELSrank) is designed to solve the MOSCC problem. Firstly, instead of use of population, two solutions (xA and xB) are generated by the greedy strategy, which have the minimal cost and the minimal time, respectively. They approximately locate in two sides of the Pareto front (PF). Secondly, with the consideration of the problem characteristics, a local search (LS) is proposed to find competitive solutions among the common neighborhood of two given solutions. If xA and xB are chosen to execute the proposed LS, solutions along the link path (the approximate PF) of xA and xB can be found. This way, the solutions along the whole PF can be found. The comparative experiments are conducted on six instances from the real-life MOSCC problems, and the results show that ELSrank performs better than other start-of-the-art algorithms, especially on the large scale problem instances.

In recent years, point cloud based data analysis has attracted lots of attentions of researchers from many different fields because of its simplicity and effectiveness. As a fundamental research, point cloud representation and recognition plays an important role. Although existing works achieved some good performances, they still cannot take full use of the information hidden in training data. This paper revisits the problem of point cloud representation and recognition from the viewpoint of data augmentation without incorporating additional data or more annotations. Different from existing works on 1D or 2D data, our proposed approach deals with a more complicated problem in three dimensional space for point cloud representation and recognition by mixing various training data from different object categories, which could help the classifier to better optimize the data-driven parameters. To validate the performance of our proposed approach, the popular used ModelNet40 dataset is employed as the standard benchmark. By carrying out comprehensive experiments under many different conditions, the experimental results show that our mixture method works positively towards improving the recognition performance of point cloud.

Point clouds provide an efficient way for 3D geometric object representation. In order to deal with the classification and segmentation of point cloud, it is very important to design an efficient and intelligent model that can directly affect point cloud. Due to the irregularity of the data format, traditional convolutional neural networks cannot be applied to point clouds processing directly. Graph convolution network (GCN) has attracted more and more attention in recent years, especially in the field of non-Euclidean data processing. Point clouds processing with GCN models is an efficient and suitable method, a lot of GCN models have achieved state-of-the-art performance on irregular data processing challenges. In this paper, we propose a Multi-scale Dynamic GCN model for point clouds classification, a Farthest Point Sampling method is applied in our model firstly to efficiently cover the entire point set, it uses different scale k-NN group method to locate on k nearest neighborhood for each central node, Edge Convolution (EdgeConv) operation is used to extract and aggregate local features between neighbor connected nodes and central node. We use ModelNet40, ModelNet10 and ShapeNet part dataset to classify point clouds and segment them semantically. Experiments show that our model achieves a better performance on classification accuracy and model complexity than other state-of-the-art models.

This paper presents an analysis of sea surface roughness from visible images based on fractals features for the first time. We measure fractal dimensions of sea surface images by the methods of box counting, fractional Brownian motion, and area measurement. The empirical values of sea surface roughness are found from measured wind speed and an empirical relation. The correlations between wind speeds (or sea surface roughness) and fractal dimensions are evaluated based on the data from a field experiment. Further, four types of noises with different parameters are introduced to the sea surface images. Then noise suppression performances of six methods are evaluated. Our experiments have demonstrated that our fractal methods and empirical relation between wind speeds and fractal dimensions are effective for analyzing roughness changes of sea surface from visible images.

A generalized memristor is introduced in Liu-Chen chaotic system for amplitude control. Different from other chaotic systems, the oscillation behavior can be rescaled in amplitude and frequency by three independent controllers. One controls the amplitude of two related variables, and the other two control both amplitude and frequency of some variables. More distinctively, the proposed memristive chaotic system shows abundant coexisting pairs of attractors under the condition of broken symmetry. The circuit experiment based on Multisim agrees with the numerical exploration.

Image super-resolution (SR) technique can improve the spatial resolution of images without upgrading the imaging system. As a result, SR promotes the development of high resolution (HR) remote sensing image applications. Many remote sensing image SR algorithms based on deep learning have been proposed recently, which can effectively improve the spatial resolution under the constraints of HR images. However, images acquired by remote sensing imaging devices typically have lower resolution. Hence, an insufficient number of HR remote sensing images are available for training deep neural networks. In view of this problem, we propose an unsupervised SR method that does not require HR remote sensing images. The proposed method introduces a generative adversarial network (GAN) that obtains SR images through the generator; then, the SR images are downsampled to train the discriminator with low resolution (LR) images. Our method outperformed several methods in terms of the quality of the obtained SR images as measured by 6 evaluation metrics, which proves the satisfactory performance of the proposed unsupervised method for improving the spatial resolution of remote sensing images.

The nonlinearity in terahertz (THz) linear frequency modulated continuous wave (LFMCW) radar usually blurs the range profile and decreases the signal to noise ratio, hampering applications where high range-resolution is particularly emphasized. A software correction method, which comprises of transmitted nonlinearity estimation and nonlinear phase compensation for the beat signal, is proposed in this paper to drastically reduce the nonlinearity in THz LFMCW radar. Besides the commonly considered nonlinearity caused by voltage-controlled oscillator (VCO), the nonlinearity from other broadband hardware devices have also been included in our modified correction model, which gives the advantage of preciser compensation. Moreover, utilizing the phase gradient autofocus (PGA) method to estimate the transmitted nonlinear term and the residual video phase (RVP) removal method to remove the range dependency of the received nonlinearity, our method can uniformly compensate the nonlinearity in the whole range profile. In addition, no presupposed parametric model for the nonlinearity waveform is needed, which further strengthens the effectiveness of the proposed method in practical use. Both the simulated data and the real tested data, acquired by a 190 GHz radar with 60 GHz bandwidth, has been used to demonstrate the validity and the effectiveness of the method.

The accurate estimation of future network traffic is a key enabler for early warning of network degradation and automated orchestration of network resources. The long short-term memory neural network (LSTM) is a popular architecture for network traffic forecasting, and has been successfully used in many applications. However, it has been observed that LSTMs suffer from limited memory capacity problems when the sequence is long. In this paper, we propose a gated dilated causal convolution based encoder-decoder (GDCC-ED) model for network traffic forecasting. The GDCC-ED learns a vector representation in the encoder from historical network traffic series, in which gated dilated causal convolutions are adopted to expand the long-range memory capacity. Moreover, different types of features in various perspectives, including temporal-independent and temporal-related features, are incorporated. In the decoder, the GDCC-ED exploits an RNN with LSTM units to map the vector representation back to a variable-length target sequence. Besides, a sequence data augmentation technique is designed to solve the problem of data scarcity. Experimental results demonstrate that our model achieves superior performance than state-of-the-art algorithms by 11.6%.

This paper proposes a novel siamese cross contrast neural network (SCCNN) to classify the hepatocellular carcinoma (HCC) and the intrahepatic cholangiocarcinoma (ICC) on computed tomography (CT) images. This method is inspired from cross contrast neural networks (CCNN) which is based on tailored CNN and information based similarity(IBS) theory. A new IBS-based measurement named as discriminative IBS(DisIBS) is designed for SCCNN. SCCNN is composed of two main parts including siamese feature extractors with DisIBS operator and MLP classifiers. Siamese networks extract features with DisIBS calculated by DisIBS operator as metric at the top. MLP classifiers are connected with but gradient-stop to feature extractors deriving classification results. We assign different loss functions with different parts to make better practice, specially DisIBS-based loss for feature extractors and softmax-based for MLP classifiers. SCCNN preserves the advantages of CCNN that can fit the insufficient medical images and small lesions. Furthermore, it extends CCNN with the siamese mechanism and gradient-stop MLP classifiers to accept the random inputs and predict like traditional CNN. To present the effectiveness of SCCNN empirically, we apply this method on a 234-person (157/77 for train/test) dataset and achieve better results than other classic CNN and CCNN methods. We try different base models of siamese structures and display prediction accuracy in two levels (slice/patient). The highest slice/patient accuracy which we have achieved on three-categories classification (HCC/ICC/Normal) is 90.22%/94.92% and the accuracy rises to 94.17%/97.44% on binary classification(HCC/ICC).

Device-to-device (D2D) communications and full-duplex (FD) communications, which can improve the spectrum efficiency (SE) of mobile communications, have received much attention. To further improve the SE, some researchers have begun to integrate FD into D2D communications, which will generate more transmission modes. Since different modes have different advantages and disadvantages, the mode selection problem needs further study. In this paper, we investigate the mode selection problem for FD-enabled two-way D2D communications to improve the SE. Four transmission modes, i.e., FD underlay mode, half-duplex (HD) underlay mode, FD overlay mode and HD overlay mode, are considered. For each mode, we maximize the system SE while fulfilling the minimum rate requirements and maximum transmit power constraints for both cellular and D2D users. For FD underlay mode, the optimization problem can be transformed into a difference of convex functions (D.C.) programming and the concave-convex procedure (CCCP) algorithm can solve the problem efficiently. For HD underlay mode and FD overlay mode, we can use search plus CCCP algorithm to solve the optimization problem. For HD overlay mode, we only need to search for the optimal solution. After obtaining the maximum SE of four modes, we can select the maximum one as the optimal mode. Numerical results are presented to illustrate the effects of the channel gains and self-interference cancellation ability on the maximum SE of four modes and the transmission mode selection.