Abstract:In order to reduce the sulfur and olefin and the loss of octane number so as to promote the clean production of gasoline,an octane number loss prediction model is established based on data accumulated by the S Zorb device.First,the Lasso is used to screen out the modeling variables,then the index factor contributions are calculated by the BP neural network,based on which 15 main variables are screened out to build the model.Second,four modeling approaches are compared and analyzed,which shows that the BP neural network has better prediction accuracy thus is more suitable to model the octane number loss.The ten-fold cross-validation produces the average MSE value of 0.027 193 and the average R² value of 0.904 87,verifying the reliability of the model.Furthermore,the main variables are optimized and adjusted by multiple linear regression under the premise that the sulfur content is not greater than 5 μg/g.The results show that multiple variables need to be adjusted simultaneously to reduce the octane number loss by more than 30%.The multiple linear regression model has good prediction accuracy and can adjust main variables positively or negatively according to a certain proportion.The trajectories of octane number and sulfur content are also visualized in the paper.

Abstract:In order to improve the classification performance of hyperspectral images with limited training samples,a hyperspectral image classification Network based on Double Pooling Attention Mechanism (DPAMN) is proposed in this paper.First,the DPAMN uses three-dimensional convolution to extract the spatial and spectral shallow information of hyperspectral images.Second,the double pooling attention mechanism is introduced into DPAMN to enhance the feature extraction ability of the network.Finally,the three-dimensional convolution dense connection module is introduced into the deep layer of the network,which can not only fully extract the spatial and spectral features of hyperspectral images,but also improve the ability of feature discrimination.Experiments show that the overall average accuracy of 95.45％,97.11％,95.30％ and 93.71％ can be achieved on datasets of Indian Pines,University of Pavia,Salinas and Houston 2013,respectively.Compared with the current mainstream advanced methods,the proposed method greatly improves classification performance on four datasets,indicating its strong generalization capacity.

Abstract:In order to prevent the mosaic line from passing through the visually significant features of orthophoto map thus damaging its integrity,a mosaic line extraction approach for orthophoto images is proposed,which combines edge information with optimized Linear Spectral Clustering (LSC) and improved A^* algorithm.First,the super-pixel segmentation theory of LSC is introduced for the mosaic line extraction,of which the classical LSC is optimized by edge intensity factor to effectively use the spectral information and edge information in orthophoto images.Second,the improved LSC is applied to the overlapping areas of two orthophoto images to obtain the boundary feature maps with various ground objects,and then the edge irregularity and isolated noise in the boundary feature maps are removed by mathematical morphology.Finally,the A^* algorithm is improved by replacing the original heuristic function based on Euclidean distance measure with Manhattan distance function,and then used to search the shortest path in the boundary map to quickly obtain the optimal image mosaic line.This approach is compared with related methods by using aerial orthophoto image taken by a UAV.The results show that the proposed approach can extract mosaic lines efficiently and with high quality,during which it effectively bypasses the visual salient features and meets the application requirements of orthophoto mapping.

Abstract:Electroencephalogram (EEG) signals are easy to record and difficult to camouflage,so EEG-based emotion recognition has attracted more and more attention.However,the diversity and individual variability of human emotion make the EEG-based emotion recognition still a difficult problem in the field of affective computing.To solve this problem,a multi-source domain adaptive dictionary learning and sparse representation approach is proposed in this study.To reduce the difference of data distribution between the source domain and the target domain,the data of all domains are projected into a shared subspace,where a common dictionary is learned.The sparse representation has the ability of class recognition according to the criteria of minimizing intra-class error and maximizing inter-class error of sparse reconstruction.In addition,each source domain adapts its domain weight to avoid negative migration.The model parameters are solved by parameter alternating optimization,and all parameters can reach the optimal solution simultaneously.The experimental results on DEAP dataset show that the proposed approach is the best among all the compared methods.

Abstract:To address the low efficiency and high false alarm rate in detection of DDoS (Distributed Denial of Service) flood attacks,this paper proposes a DWT (Discrete Wavelet Transform) and AKD (Adaptive Knowledge Distillation) self-encoder neural network based approach to detect DDoS attacks.The approach uses the DWT to extract frequency features,the auto-encoder neural network to encode and classify the features,and the AKD to compress the model in order to achieve efficient detection of DDoS attacks.The results show that the approach has high detection efficiency for proxy server attacks,database vulnerabilities & TCP flood attacks,and UDP flood attacks,with low false alarm rate.

Abstract:It has always been a difficult subject for marine radar to detect small targets on sea surface.To overcome the low detection probability of traditional detectors,a Feature Detector via Relative Sample Entropy (denoted as FD-RSE) is proposed in this paper.First,the whitened spectrum is defined to suppress the main clutter region,thus enlarge the irregularity of the sea clutter sequence.Then,by introducing sample entropy to describe the complexity of sea clutter sequence,the relative sample entropy is extracted from whitened spectrum to serve as feature.Therefore,the difference between the geometric characteristic of sea clutter and that of target echo can be thoroughly exploited in the Doppler spectrum.Finally,the superiority of the proposed FD-RSE over traditional detectors in improving detection performance can be verified by the IPIX measured dataset.

Abstract:The ground temperature changes at a rate of 0.1 ℃ every 10 years,however,a solar radiation error of about 1 K can be produced by conventional radiation shields due to the influence of solar radiation.In order to improve the accuracy of surface temperature measurement and reduce working energy consumption,this paper designs a temperature sensor based on the accelerated diffusion of radiant heat via piezoelectric ceramic vibration.First,the Computational Fluid Dynamics (CFD) method is used to calculate the radiation error of the temperature sensor under multi physical factors,then the data is fitted and analyzed using the neural network algorithm,and finally the field experiment platform is built to place the temperature sensor in real environment to verify the feasibility of the scheme.The experimental results show that the absolute error and root mean square error between the corrected value and reference value of the surface temperature sensor are 0.041 ℃ and 0.055 ℃,respectively,which also verifies the superior correction performance of the neural network algorithm.

Abstract:The non-intrusive load decomposition is to decompose the power signal of a single load device according to the known total power signal.However,deep learning based models are perplexed by problems such as insufficient load feature extraction,low decomposition accuracy,large decomposition error for infrequently used load equipment.Here,we propose an Attention Recurrent Neural Network (ARNN) model,which combines regression network and classification network to realize the non-invasive load decomposition.The model extracts the features of sequence signals through RNN network,and uses the attention mechanism to locate the position of important information in the input sequence,so as to improve the representation ability of network.Experiments on public datasets of Wiki-Energy and UK-DALE show that the proposed deep neural network is superior to the most advanced neural network under all experimental conditions.Furthermore,the attention mechanism and auxiliary classification network can correctly detect the on or off of devices,and locate the high-power signal,which improves the accuracy of load decomposition.

Abstract:It is difficult to successfully detect the false data injection attacks against the linear state estimation based on phasor measurement techniques in power systems.Here,we propose an intelligent method to detect false data injection attacks.First,the auto-encoder is used to extract the features of the power grid measurement data,which is done repeatedly to gradually reduce the feature dimension.Then the finally extracted feature is subjected to supervised learning through the Softmax layer,so as to obtain an attack detection algorithm based on stacked auto-encoders.Second,the attack detection approach is improved through noise reduction to solve the over fitting of auto-encoders.Finally,the proposed method is simulated and verified by IEEE-118 node test system,and the results show that the proposed attack detection method has high computational accuracy and efficiency.

Abstract:Aiming at the cyber physical system (CPS) subject to false data injection (FDI) attack,a control method based on sliding mode and extended observer is proposed.First,the system is dynamically linearized,an extended observer is constructed,and the convergence condition of the observation error is analyzed.Second,the integral sliding mode surface is designed,the asymptotic stability criterion of the sliding mode system is derived by using linear matrix inequality,and the sliding mode vector satisfying the gain performance of the system is obtained.Then,based on the exponential reaching law,an adaptive integral sliding mode controller is proposed to eliminate quantization errors and generalized disturbances,so that the system can reach the sliding surface.The advantages of this method include high estimation accuracy,fast response speed,and strong robustness to FDI attack and quantization parameter mismatch.Finally,numerical simulation verifies the effectiveness of the method.

Abstract:The output characteristics of photovoltaic (PV) array change with the environmental conditions and running state.In order to meet the control requirements of Maximum Power Point Tracking (MPPT) under different operation conditions,a segmented control method combining improved Quantum Particle Swarm Optimization (QPSO)and perturb and observe algorithm is proposed after analyzing the output characteristics of photovoltaic array under various working conditions.The inconsistent adaptive mutation DCWQPSO is used to search the maximum power point globally in the initial stage of tracking control to make the power point converge to the maximum power point quickly in order to improve the tracking speed,then the perturb and observe algorithm based on closed-loop fuzzy control is used to search the maximum power point locally to improve the tracking accuracy.The Matlab simulation results show that the segmented control method can complete MPPT in only 0.32 s under various working conditions of photovoltaic array and remains stable,which has faster tracking speed and higher tracking accuracy than others,indicating its capacity to improve the efficiency of PV generation effectively.

Abstract:Accuracy is an important indicator of the three-dimensional (3D) model of UAV oblique photography.Conventional modeling need to deploy Ground Control Points (GCPs),which is inefficient and costly.Control-free modeling does not require GCPs,which is efficient but not accurate.Here,we propose a 3D modeling approach for oblique photography with post-processed image control.The study was carried out in the east area of Nanjing University of Information Science & Technology.The results show that both the 3D models established by conventional modeling and post-processed image control modeling are accurate,with plane errors and elevation errors less than 0.05 m,meeting the modeling requirements of high efficiency,high precision and low cost.However,the post-processed image control modeling avoids the making and maintaining of the image control points,instead,it deploys the existing high-precision 3D model to select image control points and extract the coordinates.The proposed approach can be applied to regular updating of high-precision 3D real scene models for cities,towns,roads,etc.

Abstract:To solve the poor modeling performance caused by many blind areas and large difference in image resolution for complex buildings,a 3D surround photography and refined modeling approach is proposed for UAV photography.First,a 3D surround image acquisition route is designed for a small single-lens multi-rotor UAV according to the appearance of complex buildings.Then the acquired images are imported into Context Capture software,which are analyzed and processed for model establishment.Finally,DP-Modeler and 3ds Max are used to modify the model so as to solve the information missing such as lack of texture and fuzzy drawing.The proposed approach is then compared with traditional UAV tilt photography in modeling efficiency,model accuracy and texture details with two adjacent buildings in a park as research object.The results show that the proposed method has the advantages of high modeling efficiency,high precision and rich texture,improves the quality and efficiency of complex building modeling,solves the information missing in image acquisition via combined modification,and significantly improves the quality of model details.