3D technology is known as the trend for future TV development. Criteria of a perfect 3DTV solution includes glasses-free, narrow bandwidth, 2D-3D switchable, high definition, randomly accessible, non-intrusive interface and scalable. Two practical solutions for glasses-free 3DTV are multi-view 3D which respectively based on active lenticular, active barrier, and active backlight and active technology which relies on are active optics and eye-tracking devices.

Traditional machine learning methods and deep neural networks need a large?scale of labeled samples as support in the process of training model. However，labeling a large?scale dataset is a time?consuming process，and it is not realistic to obtain all kinds of labeled samples due to the dynamic real world. Therefore，researchers broke through the limitation of labeled samples and proposed open set recognition which is more suitable to real scenes. Open set recognition models can not only classify the categories appearing in the training process，but also deal with the unseen categories effectively. In recent years，open set recognition has developed rapidly and attracted many researchers to focus on open set problems. This paper summarizes the existing open set recognition works. First，open set recognition is defined to distinguish from other related works. Second，the open set recognition algorithms are summarized according to the model building，metric selection and property of incremental. Furthermore，two theories used in open set recognition are introduced. Finally，this paper looks forward to the future development issues of open set recognition.

 Biological soft tissue elasticity imaging using ultrasound is studied in this paper. Ordinary ultrasonic imaging use acoustic impedance as the imaging parameters, which can only reflect the change and the distribution of acoustic properties of tissue. When the lesion acoustic impedance changes very little while the elasticity varies much, it is necessary to detect the change in both the acoustic and the mechanical properties of tissue using ultrasound elasticity imaging. It has been more than twenty years since the concept of elastic imaging was put forward. In order to realize the elastic properties of the soft tissue of the real-time, fast, high resolution imaging, there has been much research work on theory and clinical experiment. Elasticity imaging gradually becomes the focus of ultrasound imaging in medical field.
The principle of ultrasonic elasticity imaging mainly includes the traditional elasticity imaging, radiation force imaging and shear wave imaging. The traditional elasticity imaging is a static compression method based on elasticity, in which the global elastic distribution of tissue can be obtained by strain estimation. The previous research mainly includes the principle, algorithm, signal processing, noise suppression and inverse problems etc. Radiation force imaging and shear wave imaging are new imaging methods based on acoustic radiation force, in which the local elastic information nearby the acoustic focusing can be obtained. Moving the focus position can obtain the global distribution of elastic properties.
In this paper, the simulation on ultrasound elasticity imaging of biological soft tissue by using finite element method was studied. By changing the internal distribution and the values of Young’s modulus, the displacement and strain distribution in different cases can be calculated and the simulation results are analyzed. In previous simulation and experimental study, soft tissue is considered as linear elastic model. In this paper, the tissue model is revised by adding consideration of biological tissue viscoelasticity. The simulation results coincide well with the real situation using this optimization of the organizational model. Simulation results show that the use of micro-strain compression method can be realized from the distribution of elastic tissue imaging, and the simulation results by considering viscoelasticity show that the stress, strain and displacement in the tissue are functions of time. 

Block effect is the noticeable discontinuous leaps in the reconstructed image, which is caused by the high frequency information of edges lost in the process of quantification using Discrete Cosine Transform (DCT). A deblocking algorithm based on Contourlet transform is proposed in this paper.This algorithm will firstly decompose the image by Contourlet Transform, then process the obtained Contourlet coefficients with the deblocking algorithm, and lastly reconstruct the image with the updated coefficients.Experiments show that the algorithm can keep more details of the original image and have better recovery performance on image edges than traditional methods. 

The noisy genetic algorithm (NGA) is a new approach to manage the spatial variation of parameters in recent years. In this study, in order to explore the application range of NGA strategy, it commenced research to study the performance of NGAstrategy under different spatial variation of hydraulic conductivity. The study results show that the optimization results of NGA strategynot only have high reliability but also have high computational efficiency when little than 1.0. However, when increased to 2.0, or even to 3.0, 5.0, the optimization results of NGA strategyno longer have high reliability. In this case, increasing the maximum sampling number of NGA can improve the accuracy of Pareto optimal solutions and reduce the uncertainty of the optimization results. However, when the maximum sampling number of NGA reach a certain extent, the accuracy and reliability of the optimization results will not significantly increase. To further improve the accuracy of the optimization results, it needs to seek other method to describe the spatial variation of hydraulic conductivity, such as increasing remediation cost to get more hydraulic conductivity condition point to reduce the uncertainty of the hydraulic conductivity. Then, more accurate and reliable optimization strategy can be found

The electromagnetic pollution has become a serious problem due to rapid development of electronic information technology. It is urgent to develop electromagnetic wave absorbing materials with "thin,light,wide and strong" features. Graphene materials have unique properties,such as high specific surface area,high electronic conductivity,low density and strong dielectric loss. However,they also have poor impedance matching and single electromagnetic loss mechanism. The poor impedance matching of graphene could be improved effectively through the regulation of the morphology,structure and so on. Besides,composite materials of combination of graphene with other electromagnetic loss materials exhibited multiple synergistic effects,which resulted high?efficiency and broadband absorption. In this paper,we briefly explain mechanism of electromagnetic absorption,review the recent development of electromagnetic wave absorbing materials and predict research areas of graphene?based electromagnetic wave absorbing materials in the future.

Dechlorane Plus(DP)is currently being used in the production and use of chlorinated flame retardants.This study established a method for analysis of DP in samples of soil，sediment and water.The method was used to survey the DP levels and distribution surrounding a DP manufacturing plant in Jiangsu Anbang electrochemical company.The concentration ranges of DP in the surface soil surrounding the plant was 0.665－102 ng·g－1 dry weight.The concentration ranges of DP in the surface sediments and water from the surrounding river were 0.165－65.1 ng·g－1 dry weight and ND?4.91 ng·L－1，respectively.The concentrations of DP in the activated sludge and the effluent water from the wastewater treatment pool were 1.23×104 ng·g－1 dry weight and 2.29×104 ng·L－1，respectively.The results of DP isomer ratios implied that syn?DP selective enrichment or anti?DP selective degradation of DP in water and sediment.It is suggested that the DP manufacturing plant is an important source of DP contamination in the surrounding environment.

It has been demonstrated that it is possible to recognize human emotions by using electroencephalogram(EEG)signals. In recent few years，the development of machine learning technology has provided reliable techniques for EEG based emotion recognition research. Traditional machine learning methods extract features from multi－channel EEG signals in each channel and concatenate them into a single feature vector，ignoring critical temporal dynamic. The Long Short－Term Memory(LSTM)in deep learning technology can solve this defect well due to the recurrent structure. However，directly fitting the LSTM based model using EEG signals consumes huge computer resources and ignores important frequency domain and nonlinear dynamic information. In this paper，we present a new emotion recognition method based on LSTM. Various features in time domain，frequency domain and nonlinear dynamic are extracted from multi－channel EEG signals and constructed into feature sequences which are used to train the LSTM based model. Experiments are carried out on the DEAP benchmarking dataset for valence，arousal and liking classification respectively，and every emotional dimension is divided into two classes(low and high). Experimental results demonstrate that the classification accuracy of the proposed model outperforms the previous methods，with regard to both of valence and liking emotional dimensions，and is also comparable to the most advanced method for arousal classification.

The establishment of an accurate battery model is critical for the estimation of SOC (state of charge) and SOH (state of health).The accuracy of model mainly depends on two aspects: the determination of the model kind and the identification of model parameters. The research shows that the second?order dynamic lithium?ion battery model can effectively simulate charging and discharging process. So this paper mainly focus on the second?order model of lithium battery. The problem of parameters identification of the model can be reduced to a nonlinear optimization problem. Firstly,a corresponding nonlinear least squares optimization model is established. Then the optimal identification results of the parameters are obtained by the Levenberg?Marquard algorithm. However the optimization problem is not only non?linear,but also non?convex. So the selection of the initial value of RC parameters is important to the outcome of the algorithm. At the same time,the variation of the discharge current will also affect the identification results. In order to solve these two problems,this paper first proposes a simple and effective method for selecting the initial value of RC parameters and a method for adjusting the objective function in the optimization model to ensure the algorithm can obtain correct solutions under different discharge current conditions. Parameters under different soc are identified by this method. The simulation results show that the algorithm can quickly and accurately obtain the RC parameter values in the model when a suitable initial value is selected.

Knowledge tracing aims to track learners' cognitive states dynamically and predict their future performance based on their historical response data. However，existing knowledge tracking models usually only utilize the knowledge concepts representing the test without considering the critical knowledge contextual features contained in the test itself，which limits the effect of the model. Moreover，compared to cognitive diagnosis methods that incorporate educational priors，the interpretability of knowledge?tracing models is inadequate. This paper proposes a knowledge context?aware deep knowledge tracing model to address these issues. The model includes a knowledge context representation module to capture deep?level knowledge weights，question difficulty，and other contextual features. In the knowledge aggregation module，the model embeds the knowledge weights into the computation of learners' abilities toward specific questions. Lastly，in the learning prediction model，the factors of guess and error are introduced，and the predictive performance in real?world scenarios is optimized through a cognitive diagnosis model to further improve the model's predictive performance. Compared to existing methods，the model proposed in this paper achieves better prediction results at the question level and demonstrates advantages in model interpretability.

Nanoparticles (NPs) have been widely used in different fields. Once NPs get in touch with organisms，a layer of protein molecules，called protein corona，will form on the NP surface quickly and NPs thus acquire new biological characteristics. Unlike the in vitro environment，the corona?containing structure is a real state of NPs in vivo. The physicochemical properties of NPs (e.g.，size，shape，and surface modification) will affect the composition of protein corona. Environmental conditions (e.g.，medium composition，incubation time，temperature，and pH) also play important roles in the formation of protein corona. On the other hand，the protein corona can affect the interactions between NPs and organisms，thus altering their biouptake，bio?distribution，and toxicity. Nevertheless，how different protein molecules specifically bind to the surface of nanoparticles is still unclear and methods to trace the dynamic changes of protein?NPs complex in living organisms are lacking. These are problems that need to be solved in the future.

Traveling Salesman Problem (TSP) is a classic problem in Combinatorial Optimization Problem (COP)，which has been repeatedly studied for many years. In recent years，Deep Reinforcement Learning（DRL）has been widely applied in driverless，industrial automation，game and other fields，showing strong decision?making and learning ability. In this paper，DRL and graph attention model are combined to solve TSP by minimizing the path length. Specifically，the behavioral network parameters are trained by an improved REINFORCE algorithm to effectively reduce the variance and prevent local optima; Positional Encoding (PE) is used to the encoding structure to make the multiple node satisfy translation invariance during the embedding process and enhance the stability of the model. Further，we combine Graph Neural Network (GNN) and Transformer architecture，and apply GNN aggregate operation processing to transformer decoding stage for the first time，which effectively capture the topological structure of the graph and the potential relationships between points. The experimental results show that the optimization effect of the model on the 100?TSP problem surpasses the current DRL?based methods and some traditional algorithms.

Montmorillonite is the most commonly distributed clay mineral in earth surface system. Exchangeability of interlayer cation is one of its characteristic properties，which makes it an important natural materials for various applications. Different cations (K⁺，Mg²⁺，Ca²⁺，Ba²⁺) exchange from montmorillonite interparticle pore fluid into clay interlayer space has been studied at atomistic level by using classical molecular dynamics simulation. The final cation exchange capacities follow the order of Ba²⁺>Ca²⁺>K⁺>Mg²⁺，as opposed to the hydration ability of cations. In the montmorillonite interlayer，Mg²⁺ is absolutely hydrated by waters and far away from the montmorillonite surface. Parts of Ca²⁺ and Ba²⁺ ions may closely hover above the tetrahedral substitution position of the clay，however，most K⁺ ions are bound by the six?membered ring of the clay surface. The migration of cation from fluid to clay interlayer space corresponds a process with decrease in free energy. The mobility of interlayer cations is much lower than that in fluid，especially the self?diffusion of Ba²⁺ is the lowest. The disclosed dynamic process of cation exchange of montmorillonite at atomistic level will enhance the understanding of clay?fluid interactions.

 Smart home system is now widely used in our houses, offices and etc. but those existing products are mostly wired，which causes inconvenient installation and high cost.Therefore we propose a wireless smart home
system based on Internet of things，which aims at expanding the range of network users from people to things，eventually making the automatic process of information transmission among things come true. In other words,
things could communicate automatically. Our smart home system has a thre-layer structure, which includes front end sensor layer, network layer and user service layer.The sensor layer consists of entrance guard，cnvmonmcnt
monitors and household-appliance controllers, all these devices support ZigBec communication protocol. Houschold- appliance controllers arc integrated into according remote controllers, in order not to modify appliance itself.The
network layer consists of EEE802. 15.4 (ZigBee)，a control terminal and global system for mobile communications (GSM)/general packet radio service(GPRS).Control terminal is in charge of information gathering and
distribution. A secure digital memory card is used to record every single event. The front end sensors connect to the control terminal over ZigBec network in a star topology.The terminal plays the role of coordinator, which is a full-
function device. And all the front end sensors arc normal reduced-function devices. Wireless ranscciver is performed by JN5139，which is a low power, low cost wireless microcontroller suitable for ZigBec.The user service layer
provides human-device interactions.Tcrminal controller supports bidirectional communication with users over GSM/ GPRS by integrating Sicmens MC35i module，which is a supper slim dual-band GSM communication module; it
supports standard attention command set. Records could be checked by using keyboard and liquid crystal display of terminal.The whole smart home system functions in this way; once front end sensors detect anything unusual，they
would notify the control terminal immediately over ZigBec network. After receiving a notification, the control terminal would parse it，and send a text or multimedia message to user’s cell phone over GSM/GPRS network if
necessary, In the opposite direction, if user wants to control a household appliance, the only thing he needs to do is to send a formatted text message to control terminal. After receiving and parsing the message，the control terminal
understands user’s intention correctly.Then terminal would send commands to corresponding household-appliance controller over ZigBec network. All communications in our smart home system are wireless，so there’re no wires
needed during installation. Human-device interactions arc mostly done over the GSM network, which has the widest coverage in this country. This makes sure our system is able to provide higlrquality service even in some remote regions.

Carbon dioxide pipeline transportation technology is the link between carbon dioxide capture technology and utilization technology. It connects the origin and storage places. It can continuously transport carbon dioxide with high economic benefits and high cost performance,which is in line with the criteria of sustainable development. Among them,supercritical transportation is the main way of carbon dioxide pipeline transportation in the future. This paper systematically and concretely introduces the development of CO₂ pipeline transportation from four aspects: the principle of pipeline transportation,the current situation of CO₂ pipeline transportation technology at home and abroad,the research on safe transportation control technology of CO₂ pipeline at home and abroad and the typical CO₂ pipeline transportation demonstration project,and puts forward the trend prospect of future development.

The soils on the lunar surface represent the boundary between the solid Moon and the solar system,which play an important role in understanding the Moon and its surrounding space environment. Since the return of lunar samples during Apollo era,lunar soils with complicated origins and diversified components have been classified to facilitate research with different aims. It becomes necessary to carry out such a review with the return of CE?5 lunar soils. By summarizing classification schemes of lunar soils in the literature,this paper discusses the relevant research directions and achievements using two different classifications: particle size and component (grain). New classification methods and applications have also been briefly discussed. This review may shed light on the future research of lunar soil samples.

B3LYP/6?31++G(d,p) and M06/6?311+G(d,p) methods were applied to investigate the coordination ability of glycine and the first series transition metals. In the most stable structure of every kind metal complex,coordination atoms from glycine are the two oxygen atoms. The order of coordination ability of each metal from big to small is as follows: Cu²⁺>Ni²⁺> Co²⁺>Zn²⁺>Cr²⁺> Fe²⁺>V²⁺>Mn²⁺>Ti²⁺>Sc²⁺,and binding energy of the most stable complex is -1012.1 kJ?mol^-1. In each structure,the rate of orbital interaction energy and electrostatic energy is close to 50% respectively,The dispersion energy is only 0.2%. The more orbital interaction energy ,the more the number of electrons transferred.

For a decade，sharing information through social networks (e.g. Microblog and Twitter) has become an indispensable part in our daily life. Therefore，how to effectively predict the social influence has become an important subject in the study of social network. There are many applications，such as identifying viral marketing and fake news,accurate recommendations and online advertising. In recent years，some deep learning social network influence prediction methods have made some progress，but still faces the following difficulties: users typically have different behaviors and interests and they interact through different channels at the same time,and relationships between users are difficult to detect and formally describe. Traditional social network influence prediction methods manually extract the feature information of users and their networks by designing complex rules. However，the effectiveness of this method heavily depends on the domain knowledge of the rules set，which makes it difficult to generalize the rules in one field to the application in other fields. Based on deep learning method，we design an end?to?end neural network to learn the features of users' hidden information to predict their influence in the social network. Firstly，feature extraction is carried out on users' local network through graph embedding，and then the graph neural network is trained with feature vector as input，so as to predict users' social representation. Compared with previous work，we combine the feature attributes of users in social network with the local area network features by using graph convolution and graph attention method，which greatly improves the accuracy of model prediction.

With the continuous development of the Internet technology，network security threat intelligence analysis that base on security knowledge graph(SKG) can analyze multi－source threat intelligence data in a fine－grained manner，which has received extensive attention. Traditional named entity recognition(NER) methods are difficult to identify network security entity which mix Chinese and English in the field of network security，and can't fully extract some features，so it is difficult to accurately identify the network security entity. In this paper，we propose a novel CNN－BiLSTM－CRF security entity recognition method combining with feature template(FT－CNN－BiLSTM－CRF)on the basis of deep learning model. The feature template is used to extract local context features，and neural network model is used to automatically extract character features and text global features. Firstly，each character of the input sequence is converted into a corresponding character vector，and the convolutional neural network(CNN) extracts the character－level features. Secondly，the character－level features vectors are input into the BiLSTM(Bi－Long Short－Term Memory) together with the local context vectors extracted by the feature template. The global features of the security entity are automatically extracted by BiLSTM. Finally，the CRF(Conditional Random Fields) labels the network security entity to obtain the recognition result of the security entity. The experimental results show that our method reaches 86% F－scores on the large－scale network security dataset and outperforms other methods.

Clustering algorithms are commonly difficult to determine parameters and to adapt to datasets of various shapes. And their time complexity increses with theire improvement of universality. To solve these problems,we propose a new clustering algorithm which finds the mutation position of the sample point distance by combining the global and the local features，combines the global and the local features look for the mutation position of the sample point distance，and realizes the clustering of the convex spherical datasets by calculating the minimum distance within the cluster of the sample point. On this basis,we propose the concept of sub?cluster connectivity which adapts to datasets of various shapes by merging sub?clusters based on two easily determinded parameters. Comparing this algorithm with DBSCAN （Density?Based Spatial Clustering of Applications with Noise） and other clustering algorithms on four classic datasets，the experimental results show that the proposed algorithm can be applied to datasets with complex cluster shapes，and has faster calculation speed than algorithms with the same clustering ability. Its advantages include fewer parameters which are easier determined with excellent comprehensive performance.

Today when network technology is very mature,all kinds of sensitive words including pornographic,violent,politically words are flooding websites. The detection and identification of these words is necessary to create a healthy network environment. Most of these sensitive words attempt to camouflage through the pronunciation or glyph to deceive the detection system. The existing matching algorithm can detect words with exactly the same pronunciation,but can not accurately identify the variants with similar pronunciation or similar fonts. To solve this problem,a Chinese character similarity comparison algorithm for fuzzy matching is proposed here. Firstly,through the special coding of Chinese characters,this paper proposes an improved algorithm for the similarity of Chinese characters in sound and shape codes considering the characteristics of pronunciation and font. Then,with the traditional dictionary tree,precision parameters are added to set the matching precision,so as to complete the sensitive word detection. The experimental results show that the matching accuracy is improved by 8%~39% and the error rate is reduced by 6%~38% on the commonly used similar Chinese character dataset.

Electrocaloric effect (ECE) leads to a novel technology of refrigeration and heat pumps utilizing condensed matter. The ECE is originated from reversible manipulation of dipolar ording during the electric field induced strutrual phase transtion of polar dielectrics. Due to the nature of a field?effect,the electrocaloric (EC) refrigeration cycle exhibit ultra?high theoretical coefficient of performance (COP),high cooling power density as well as device scalibitity,ease of maintaining,low?noise,and etc. In all reported EC materials,EC polymers are unique for their low dielectric and conduction loss,which exhibit 85% of energy recovery during the charging?discharging cycle. The EC refrigeration directly couples with the power line of electricity,without secondary energy transduction,which grants the technology a great potential working with household and industrial environment. The integration of these advantages over the conventional technologies was recognized as the promising next?generation refrgieration by international agencies. However,the recent development of EC devices are being still limited by the drawbacks that are inherent in the monolithic EC materials such as ceramics and polymers. Therefore it is of great significance to design and fabricate EC nanocomposites that combine advantages of individual materials. Here we will review the recent development and advances of EC nanocomposites and their bearing potential in flexible refrigeration and heat pumps. We will then look forward to the future development of EC technolgy and its contribution in zero?carbon technologies.

The rapid development of microservice technology provides technical support for enterprise system integration and network business collaboration. The process in software service system needs to evolve continuously to adapt to the needs of business changes. Existing researches mostly assess the impact of service process evolution from a single dimension. This paper proposes a two?layer dependency model，DoubleDM，for the service system including the process layer and the service layer to analyze service system evolution from two aspects: service evolution and process evolution. Regarding to the evolution of the service layer and the analysis of impact range of service modification based on the dependency between services，the expression of service dependency，the influence scope of dependency evolution and the solution with corresponding algorithm are given. For process layer evolution with various types，the process simplification algorithms on the foundation of process dependency is given. Finally，the implementation logic of the process evolution in the microservice system is proved，and the analysis is carried out by taking the example of microservice system to deal with the supply chain process.

In recent years，with rapid development of the global economy，more and more investors participate in financial investment. How to automatically choose trading strategies in complex financial markets to maximize returns has also become a research hotspot. Reinforcement learning finds the optimal trading strategy through interaction with the actual environment，so as to maximize the return on investment. Most of the existing methods apply one or two reinforcement learning algorithms to the financial market，and compare the performance of the algorithms in a single trading task. In addition，most of these studies aim at foreign stocks，securities market or cryptocurrency market，and there is little research on domestic financial market. Aiming at the above problems，this paper systematically verifies the effectiveness of different types of deep reinforcement learning representative algorithms in three investment tasks: single stock trading，multiple stock trading and portfolio allocation. The algorithm is compared by observing the back test results on the evaluation indicators such as cumulative yield，sharp ratio and maximum retracement. The results show that the appropriate reinforcement learning algorithm in different investment tasks effectively improve the income.

The capacity and internal resistance are important indicators to estimate the state of health (SOH) and predict remaining useful life (RUL) of lithium?ion batteries. However,the capacity and internal resistance of lithium?ion batteries are difficult to be directly measured online. In this paper,two health indicators are extracted after analyzing the characteristics of charging current and voltage changes during the charging process of lithium?ion batteries. Through analysis,it is concluded that the indicators are highly correlated with the battery capacity,and a two?indicators linear regression model is established to estimate the battery capacity. On this basis,BP neural network and particle swarm optimization are combined to design the SOH estimation algorithm of lithium?ion batteries. Considering that there is a certain mapping relationship between SOH and RUL of lithium batteries,the RUL prediction algorithm of lithium?ion batteries is designed by using the health indicators and the SOH estimation results. The experimental results show that the proposed indicators can accurately estimate the battery capacity and can be applied to online SOH estimation and RUL prediction of lithium?ion batteries.

We applied two organic matter removal methods to determine the effect of organic matter(OM)on the characterization and calculation of structural details of illite/smectite mixed layer mineral using XRD. A similar XRD pattern was observed between UT(untreated sample)and HT(OM removal with hydrogen peroxide(H2O2－30%)). In contrast，NT(OM removal with disodium peroxodisulfate(Na2S2O8))showed a different XRD pattern including peaks at 15 (AD，Air－dried) and 16.7 (EG，Ethylene glycol－saturated) which were not observed in the XRD patterns of UT and HT. In addition，after OM removal with Na2S2O8 treatment，the d001|001 and d002|003 peaks of I/S shifted to a higher 2－Theta angle and its d001|002 peak shifted to a lower angle，denoting an improved expandability both in AD and EG condition. A multi－specimen fitting model applied to their AD and EG patterns also suggested a much higher proportion of expandable layers after Na2S2O8 treatment. This inconsistency could be explained by the occurrence of interlayer OM in smectite layers rendering them hydrophobic，thus blocking the intercalation of water molecules and polar ethylene glycol molecules. This highly proscribes the expandability of parts of smectite layers in the AD and EG conditions and leads to an underestimate of the proportion of smectite layers in I/S. Our results clearly show that，compared to H2O2，Na2S2O8 could be more efficient in removing the inhibiting effect of interlayer OM on smectite expandability when exposed to water and polar organic molecules，which would guarantee a more precise characterization and calculation of structural details of I/S using XRD.

Lidars are widely used in the remote sensing of clouds,aerosols,atmospheric composition,temperature and wind speed. A great variety of lidars have been applied in atmospheric sounding,and many different methods has been used to classify these lidars in practice. A specific lidar can be named with different classification methods,and lidars in a specific category could be distinct,so it is useful to summarize the classification methods for atmospheric lidars.In this paper,atmospheric lidars are sorted out using more than 8 classification methods,and the features,advantages,and disadvantages of each lidar class are described briefly. According to the observation angle,atmospheric lidars can be divided into backscatter lidars and bistatic side?viewing imaging lidars (i.e.,Clidar). When the spectral features of sensors are used as the primary classification criterion,atmospheric lidars can be categorized as ordinary elastic lidars,differential absorption lidars (DIAL),Raman lidars,High Spectral Resolution Lidars (HSRL),Doppler lidars,and fluorescent lidars. According to the primary detecting target,the lidars can be named as ceilometer lidars,cloud and aerosol lidars,CO₂ lidars,ozone lidars,wind lidars,and visibility lidars. According to the lidar platform,there are ground?based lidars,airborne lidars,and spaceborne lidars. Based on the properties of light transmitter,we have solid?state lidars,gas?state lidars and semiconductor Lidars. According to the wavelength of the transmitted laser,there are ultraviolet lidars,visible light lidars,infrared lidars,and multi?wavelength lidars. By the wave mode of transmitted laser,lidars can be grouped into pulsed?wave lidars and continuous wave (CW) lidars. Atmospheric lidars can also be considered as polarization lidars (linear polarization or circular polarization) and non?polarization lidars by the ability to measure the depolarization of backscatters,classified as one?dimensional,two?dimensional or three dimensional lidars by the detection space,categorized as single?line lidars and multiple?line lidars by the number of transmitted beams,or classified as single field of view lidars and multiple field of view (MFOV) lidars based on the number of field of views. In the future,more types of atmospheric lidars would occur in response to the development of lidar techniques.The relationships between different lidar categories are discussed in the paper. For example,Raman lidars,which are typically designed to be backscatter lidars with depolarization measurements,are frequently used in the detection of cloud and aerosol lidars. Differential absorption lidars,which are frequently designed to be backscatter lidars without depolarization measurements,is usually used in the measurements of gas concentration,and they can be designed be either ground?based or spaceborne lidars.This paper might be useful for the selection of lidars in atmospheric measurements. In practice,the type of lidar should be chosen based on the advantages and disadvantages of each lidar category according to the detection target,data quality requirements and budget.

Deep learning based on neural network has made remarkable progress in wide domains. However，the latest research shows that deep learning bears the risks of privacy leakage. The current defense mechanisms are limited by the assumption of the adversary's background knowledge，the high complexity and the non?universality. This paper aims to construct a novel defense mechanism for deep learning using differential privacy. At present，the widely used differential privacy algorithm is DPSGD (Stochastic Gradient Descent with Differential Privacy) in deep learning. However，parameter setting is difficult for DPSGD，and the measurement of privacy loss is also complex. We propose DPADAM (Adaptive Moment Estimation with Differential Privacy) as a new deep learning optimization algorithm with privacy protection，which combines Adam gradient optimization algorithm and differential privacy. Also，we introduce zCDP (Zero?Centralized Differential Privacy) as a measure of privacy loss，which is more flexible and accurate. Extensive evaluation results show that DPADAM can reduce dependence on parameter settings and improve the model's fitting effect.

Concept drift is one of the main characteristics of stream data. How to detect the occurrence of concept drift and adjust the prediction model to adapt to the phenomenon of concept drift has attracted attention of researchers. At present，most algorithms about concept drift only aim at single type of concept drift detection，and need to restrict input data to obey a certain distribution，which makes the effect of detecting multiple types of concept drift unsatisfactory. An online ensemble adaptive algorithm (KSHPR) is proposed，which is optimized and improved on the basis of Adaptive Random Forest (ARF) and Streaming Random Patch (SRP) algorithms，adopting the strategy of combining non?parametric test and sliding window for concept drift detection，and reducing the influence of the window average on the performance of the algorithm. Based on this，we establish an ensemble learning model of four basic learners，which dynamically allocates weights according to the prediction accuracy of the basic learners，and effectively solves the problem of low accuracy of the learning model in streaming data. Experimental results show that the proposed algorithm performs well in both real datasets and synthetic datasets. Compared with other algorithms，the stability，classification accuracy and multi?type concept drift adaptability of the algorithm are improved.

The rapid development of smart cities has brought great convenience to people's daily lives. Among them，the increasingly intelligent video surveillance system is the inevitable result of the gradual maturity of information technology. Human behavior recognition is one of the important tasks in the field of intelligent security monitoring. However，a large number of edge monitoring devices have produced blowout image and video data. The traditional single?cloud computing model has been unable to effectively deal with the calculation and processing of massive data. This paper proposes a human behavior recognition mechanism that uses edge?cloud collaborative computing driven by big data，which expands the previous centralized computing to edge and cloud collaborative processing. Firstly，at the edge node N_{\mathrm{0}}，the video is preprocessed to remove similar frames and the extracted skeleton sequence is expressed in multiple levels. Then，the cloud trains the Spatial Temporal Graph ConvNet (ST?GCN) model and deploys it to the edge nodes N_{\mathrm{1}}~N_m. And the Edge uses the trained model to complete behavior recognition tasks and uploads the results to the cloud for fusion to obtain the final behavior category. The experimental results prove that the proposd algorithm effectively reduces the network transmission volume and cloud storage pressure problems of the previous centralized computing. And the advantages of edge?cloud collaboration make the model recognition accuracy rate steadily increasing more than 2.2%.

The tensor decomposition and deep learning have been applied to the recommendation systems and better results have become true. The tensor decomposition algorithm better extracts the hidden features of the users，recommended objects and other influencing factors from the user rating data，and the features are matched each other to give recommendation strategies. But the algorithm ignores the features in the auxiliary data information of the user，recommended objects and other influencing factors. Deep learning extracts the features of users，recommended objects and other influencing factors from auxiliary information，and matches them to give recommendation strategies，but ignores the implicit characteristics of users，recommended objects and other influencing factors in user rating data. A blending recommendation algorithm based on tensor decomposition and deep learning is proposed，which blends the two recommendation methods of tensor decomposition and deep learning. Tensor decomposition algorithm and deep learning are used to extract user features and recommendation object features from third?order user rating data and multi?source heterogeneous auxiliary information respectively，and then match them to obtain prediction ratings of user's demand or preference for recommendation objects，and the prediction ratings from the two algorithms are blended to give the final comprehensive ratings. The blending recommendation algorithm will improve the accuracy of personalized recommendation. Compared with traditional collaborative filtering algorithm，the error of blending recommendation algorithm is reduced by 34.0%.

The manufacturing industry produces massive multi?source heterogeneous data such as texts，images，audio and video in the process of design，production，sales and service. The major problem facing manufacturing companies is how to efficiently manage and use these data resources to create value for manufacturing reproduction. Traditional data storage and retrieval systems classify these data according to different forms or modalities and process them separately，resulting in a lack of correlation between cross?modal data (texts，images，audio and video data cannot be checked each other). It cannot support the problem of manufacturing business processes. In this paper，we design and implement an efficient and fast cross?modal retrieval system for multi?source heterogeneous data such as texts and pictures to realize efficient management and retrieval of multimodal data. Specifically，the system projects the these data into a unified high?dimensional semantic space for representation，generates semantic vectors，and stores the multi?source heterogeneous data in different modes according to different query requirements. Then，the system designs an efficient retrieval method of three?level structure + layered Unicom naive composition algorithm，and indexes the multimodal data according to the semantic vector to meet the semantic query needs of manufacturing users. We conduct experiments on the flickr30k dataset. Experimental results show that: (1) This system can support millions of data storage and retrieval. (2) With millions level data，the system retrieval rate is milliseconds. (3) The retrieval accuracy is higher than traditional vector retrieval methods.

Ultra?wideband low noise amplifier is essential for reconfigurable RF front?end and wideband phased array radar，and also is a dominating influence on sensitivity of the systems. This paper demonstrates an ultra?wideband low noise amplifier with integrated temperature compensation active bias circuit. The temperature compensation active bias applying in the LNA (Low Noise Amplifier) can effectively reduce the gain fluctuation caused by variation of ambient temperature. Meanwhile，the bandwidth expansion technology is utilized to improve the high frequency gain of transistor and realize the 9?octave operating bandwidth. The proposed LNA is fabricated by using 0.15 μm GaAs pHEMT technology with a dimension of 2 mm×1.2 mm. The measurement results show that the LNA has low power consumption of 125 mW with 5 V operating voltage. The operating bandwidth is 1~9 GHz with noise figure of less than 1 dB and gain of above 25 dB.The input and output return loss is less than -10 dB，and the output 1 dB compression point is above 10 dBm. Significantly，the gain fluctuation of the LNA is less than 1 dB in ambient environment of -55~125 ℃.

The popularization of the Internet has enabled the rapid development of online education，which has not only alleviated the imbalance of educational resources，but also provided sufficient educational data for researchers. As an emerging discipline，Educational Data Mining (EDM) aims to understand students' learning behaviors and provide personalized suggestions by analyzing the educational data. Knowledge tracing is an important task in EDM，which models students' historical answer sequences to predict their next answer performance. Existing knowledge tracking models do not distinguish long?term and recent interaction information in historical sequences，and ignore the sequence information at different time scales on future predictions. In this paper，we propose a novel knowledge tracing method based on multi?scale attention fusion，which uses temporal convolution networks to capture multi?scale information of historical interaction sequences，and performs multi?scale information fusion based on the attention mechanism. For different students and historical sequences，the model can adaptively determine the importance of different time scales. Experimental results show that the performance of our model is better than the existing knowledge tracing models.

The key to the neighborhood search algorithm is the selection of neighborhood structure，but the search time of each iteration is long，and the ability to search autonomously in the solution space is lacking. In this paper，the deep reinforcement learning (DRL) model is used to improve the neighborhood search algorithm，and a new deep hybrid neighborhood search (DHNS) model is designed to solve the capacitated vehicle routing problem (CVRP). Firstly，the greedy algorithm is used to provide the initial solution for the DRL model. Secondly，the pointer network and Transformer hybrid encoder are used to take advantage of different networks to extract node feature information at a deep level. Finally，the repair process of the repair operator is transferred to the DHNS model，and the process of neighborhood search repair solution is automatically completed，expanding the ability to solution space for autonomous search. At the same time，aiming at the complex transmission mechanism in hybrid encoder and the problem of decoding misleading information，the AOA （Attention on Attention） mechanism is further added in the encoding and decoding process. AOA is responsible for screening valuable information，filtering out irrelevant or misleading information，effectively characterizing the correlation between attention results and queries，and modeling the relationship among nodes. Experimental results show that the DHNS model is superior to the existing DRL model and some traditional algorithms in the optimization effect of 100?scale CVRP. The efficiency of the algorithm is verified by using an example in the CVRPlib dataset，and the results show that the DHNS model greatly improve the optimization efficiency of the routing problem.

The optical and electrical properties of phosphorus doped silicon carbide thin films with various C/Si ratio were studied before and after thermally annealing.It is found that with decreasing the C/Si ratio for as­deposited samples，the optical band gap is gradually decreased and dark conductivity increased accordingly.As high as 6 to 7 orders of magnitude of material dark conductivity improvement is a remarkable result of 1000 ℃ annealing.With decreasing the C/Si ratio for annealed samples，the Si－C bond density is decreased in company with the enhancement of crystal degree and optical band gap，and also with improvements of main carrier mobility as well as dark conductivity.Besides，influences of the C/Si ratio to the phosphorus dopant activation effect and to the material conductivity activation energy are also significant.The dopant activation effect changes in the form of the carrier concentration，which increases firstly and then decreases slightly with reducing the C/Si ratio，representing a close relation to the crystal degree.Furthermore，the conductivity activation energy of annealed samples is reduced with decreasing the C/Si ratio，the Fermi level consequently rises very close to the bottom of the conduction band or even enters into the conduction band，indicating the formation of heavily doped semiconductor materials.

In recent years，soil microplastics pollution is becoming more and more serious，therefore it is urgent to focus on its ecological effect. In this study, we took Vicia faba (V. faba) as the test plant and treated it with polyethylene microplastics (PE?MPs) of different concentrations (0，10，100，500 mg·kg^-1) and 50~100 μm size in soil for 30 days. By measuring the biomass，photosynthesis rate，chlorophyll content and fluorescence parameters of V. faba and the activity of antioxidant enzyme in root，the toxic effect of PE?MPs on seedling was studied. The results showed that PE?MPs could promote the growth，material accumulation and photosynthesis of V. faba at low concentration (10 mg·kg^-1) however higher concentrations (100，500 mg·kg^-1) had significant inhibitory effect (p<0.05). The superoxide dismutase (SOD)，peroxidase (POD) and catalase (CAT) activity of V. faba root increased compared with control group，but the change of malonic dialdehyde (MDA) content was not obvious. The change of soil cation exchange capacity (CEC) and the content of main nutrient elements showed that PE?MPs had a certain improvement in the ability of soil to adsorb ions and other nutrient elements. The above results provide basic data and scientific basis for studying the effect of microplastics on soil ecosystem, and are of great significance to the potential toxicity effect of higher plants and the risk assessment of food safety.

Deep learning requires a large number of labeled training samples，however，it is expensive or even impossible to obtain a large number of labeled data at first in real world. In this paper，a new deep active recognition framework is proposed，which is based on the cognitive process of how we learning and acquiring new knowledge step by step. On the basis of analyzing the cognitive errors of the model based on a small number of samples，the transformation of cognitive errors is defined，and the corresponding knowledge is obtained to actively enhance the cognitive information of the model. Based on cognitive knowledge，sensitive samples are selected to fine tune the model online. At the same time，in order to avoid forgetting the previously learned knowledge，the previous training samples are selected as refresh samples. Experiments on general datasets show that the Target Sensitive Samples (TSSs) can improve the performamce of target recognition，and the proposed cognitive learning mechanism can effectively improve the performance of deep model. The characterization of cognitive information effectively restrains the interference of other samples on the model cognition，and the online training method significantly saves a lot of training time，which provides an effective application of cognitive learning in the situation of a small number of data samples.

Motor imagery (MI) intention recognition based on electroencephalography (EEG) signals is an important issue in brain?computer interface (BCI) research. However，EEG signals have serious individual differences，while the spatial distribution of EEG signal characteristics between different subjects is very different，and the classification model between different subjects cannot be universal. To solve this problem，this paper proposes a weighted logistic regression transfer learning method based on Euclidean space.The algorithm first aligns the EEG data of different subjects in Euclidean space to make the signals more similar and reduce the difference，and calculate the different feature values?obtained by the common spatial pattern (CSP) of a specific subject. Then，it calculates the KL (Kullback?Leibler) divergence of these eigenvalues，using the KL divergence to adjust the weighted logistic regression algorithm of transfer learning to obtain the classification model.The experimental results show that， for the data set 2a in the BCI competition IV，the proposed method greatly improves the learning performance of BCI，and the classification accuracy is 15% higher than that of the baseline algorithm (Linear Discriminant Analysis）.In case of an increase in data samples，the classification accuracy of the subjects has been significantly improved. Compared with similar algorithms，the classification accuracy of the proposed algorithm increases by 4%，indicating that the algorithm in this paper furtherly improves the BCI learning performance and the generality of the classification model.

Aiming at removing irrelevant and redundant attributes，this paper proposes a method of approximate reduction to better require decision rules in intuitionistic fuzzy decision tables. The weighted system is defined via the introduction of a weighted score function in intuitionistic fuzzy sets. Additionally，variable precision rough sets are utilized for a better tolerance to misclassify. Hence，the weighted variable precision intuitionistic fuzzy sets are defined. On the basis of the constructed system，we give conceptions of the judgment theorem and identification matrix of both lower and upper approximate reduction，by which two approaches of reduction are put forward. Finally，a concrete example and numerical tests are used to illustrate the effectiveness of the proposed method.