Computer and Modernization

Water Level Prediction Based on SFLA-CNN and LSTM Combined Model

ZHOU Yong-qiang, ZHU Yue-long

2021, 0(04): 1-7.

Asbtract ( 280 )

PDF (1152KB) ( 253 )

References | Related Articles | Metrics

Hydrological time series are affected by rainfall, showing seasonal and non-linear characteristics. The traditional single model is simple in structure, which has the problems of low prediction accuracy for complex nonlinear hydrological time series and can not capture the composite characteristics of hydrological time series well. The combined forecasting model adopts the idea of multi-classifier, which can effectively improve the forecasting accuracy. However, it is time-consuming and inaccurate to manually adjust the model parameters. In this paper, a combined forecasting model based on SFLA-CNN and LSTM is proposed: The parameters of CNN model are optimized by Shuffled Frog Leaping Algorithm, and the optimized SFLA-CNN forecasting model is obtained; After that, BP neural network is used to nonlinearly combine the predicted values of SFLA-CNN and LSTM models to obtain the final prediction results. The experimental results of water level prediction in Taihu Lake region of Jiangsu Province show that the combined model has effectively improved the prediction accuracy and better generalization ability compared with the existing models.

Breast Cancer Diagnosis Model Based on BP-GamysBoost

LIU Jun, PENG Hui-xian, HUANG Bin, Tony SHAY

2021, 0(04): 8-14.

Asbtract ( 150 )

PDF (1459KB) ( 141 )

References | Related Articles | Metrics

In view of the problem of unbalance of the breast cancer data, the standard Adaboost algorithm is improved. First, BP neural network is introduced, then the strong global optimization ability and fast convergence speed of simulated annealing genetic algorithm (SA-GA) are fused, and finally the weight is allocated reasonably to propose the BP-GamysBoost algorithm. At the same time, in order to verify the rationality of the proposed new algorithm BP-GamysBoost, the WBCD database is obtained from the UCI machine learning knowledge base, and the performance indexes such as stability, accuracy, missed diagnosis rate and sensitivity of BP-GamysBoost algorithm model are compared with BP model, BP-GA model and BP-Adaboost model. The results show that the BP-GamysBoost model works well in the breast cancer database and is superior to the other three algorithm models.

Research on Depth of Oil Well Moving Liquid Surface Based on Short-term Energy and LSTM

LIANG Xin, ZHANG Zhu-hong,

2021, 0(04): 15-19.

Asbtract ( 169 )

PDF (744KB) ( 148 )

References | Related Articles | Metrics

Dynamic oil well liquid surface depth estimation has been being a crucial issue in the field of oil. It will be extremely important for the development of oil enterprise how to efficiently and precisely acquire the dynamic information of liquid surface depth. Therefore, for the problem that the depth estimation accuracy of oil wells’ dynamic fluid surface is influenced greatly by environmental noises and depth estimation errors, the current work probes into the algorithms of oil wells’ surface depth estimation and prediction based on acoustic wave curves. Therein, a depth estimation algorithm suitable for estimating the depth of dynamic liquid level is acquired through designing an improved short-time energy zero-crossing function and an improved three-electric center clipping function, in which multi-channel liquid level position estimations are fused to decide the position of liquid level. After that, a liquid surface depth prediction model is obtained based on the LSTM neural network, in which the gained liquid surface positions and average sound velocity are taken as the input of the network, and the actual liquid level depth is viewed as the desired output. The comparative experiments have confirmed that the current depth estimation method can effectively decide the depth of dynamic liquid surface, and the prediction model can well predict oil wells’ liquid surface depth.

Analysis of Emotional Tendency of MOOC User Comments Based on BERT and Bidirectional GRU Model#br#

Nigara Masimjan, Azragul Yusuf

2021, 0(04): 20-26.

Asbtract ( 434 )

PDF (1007KB) ( 154 )

References | Related Articles | Metrics

For the purpose of realizing the sentimental analysis of the user reviews of MOOC, this article proposes a method for the sentimental classification of user reviews based on the BERT and BiGRU model. The article uses the BERT model to extract the feature representation of the course review text, and the acquired words features are input to the BiGRU network to extract the emotional features of user reviews, finally the emotional tendency classification is performed by Softmax logistic regression. Experimental results show that the F1 value of the review sentiment orientation classification model based on BERT and BiGRU model reaches 92.45%, which improves the accuracy of user sentiment orientation analysis and is better than other mainstream orientation analysis models, proving the effectiveness of the method.

Rural Novel Translation Methods Based on Bidirectional GRU Neural Network Machine Model

SUN Li-li, GUO Lin, WEN Xu, ZHANG Wen-nuo

2021, 0(04): 27-31.

Asbtract ( 187 )

PDF (1112KB) ( 110 )

References | Related Articles | Metrics

In order to improve the accuracy and efficiency of machine translation of novels, an RNN neural network framework based on end to end is proposed to study the rural novel translation methods by using the Chinese to English machine model. By analyzing the translation principles and performance of the RNN-NMT model, WordNet Semantic Similarity model, GRU-LM model and BiGRU-LM model, the new BiGRU-LM-Attention machine model is established to carry out translation testing and quality performance evaluation. Tests prove that the BLEU value of the new model is higher than these of other models; at the same time, on the quality performance of example translation, the accuracy rate of new model is ahead of 4 online translating tools in terms of semantic recognition, dialects, special nouns, slang and flexible recognition of passive voice, indicating that the improved neural machine model in this paper can adapt to the characteristic language of the novel and effectively improve the translation quality, which is of significance of Chinese culture transmission.

Ensemble Stacked Autoencoders and XGBoost Based Deep Learning Model for Significant Wave Height Forecasting

LU Xiao-min, LIU Fan, CAI Li-hua, LI Xue-ding, XU Xiao

2021, 0(04): 32-36.

Asbtract ( 180 )

PDF (1103KB) ( 85 )

References | Related Articles | Metrics

The significant wave height forecast is crucial for both human marine activities and marine engineering. The artificial neural network has been widely used in significant wave height prediction and achieved good results. However, as a shallow network architecture, it has limited expressive ability, making the forecast accuracy fluctuate in different regions. Therefore, to improve the overall forecast accuracy of the significant wave height, this paper proposes a deep learning model of significant wave height forecasting by integrating stacked autoencoders (SAE) and XGBoost. First, the powerful feature representation capabilities of the SAE algorithm are used to process ocean wave data to realize the extended dimension expression of the data. Secondly, the deep feature expression of SAE is used as the input of the XGBoost algorithm to predict effective wave heights. This paper focuses on the significant wave height prediction method and uses the measured wave data of Buoy 2 in the central Taiwan Strait in 2017. The experimental results show that our approach is superior to existing methods in terms of deterministic coefficient (R^2) and mean square error (MSE).

An Analysis Method of Influencing Factors of Legal Judgment Prediction

YIN Min, LI Xiao-hui, LI Chang-bao, GU Ping-li, ZHANG Ke, LYU Shou-ye

2021, 0(04): 37-41.

Asbtract ( 171 )

PDF (829KB) ( 80 )

References | Related Articles | Metrics

Legal judgment prediction refers to the method by which machines predict the crimes of a legal case based on the description of the facts of the case. It is a promising application of artificial intelligence technology in the legal field. With the rapid development of the field of artificial intelligence, the application of related technologies in the legal field has become more and more extensive, and many classic models have achieved good results in the prediction of legal judgments. Although the effect of machine learning methods is good, it has been unable to solve the problem of interpretation and reasoning of the prediction results. The prediction results have black-box characteristics, and the supporting basis for the conclusion cannot be obtained. In response to the above problems, this article proposes an analysis method of influencing factors of legal judgment prediction, which combines Chinese word segmentation technology, support vector machine technology (Support Vector Machine, SVM) and unified interpretation and prediction framework (SHapley Additive exPlanations, SHAP) to achieve intelligent prediction of judgment results of legal cases, scientific analysis of the influencing factors of the prediction results, and key factors that have a greater impact on the prediction results, providing a supporting basis for the prediction results.

Damaged Old Photos Inpainting Based on Generative Adversarial Networks

CHEN Yuan-yuan, LIU Hui-yi

2021, 0(04): 42-47.

Asbtract ( 241 )

PDF (5368KB) ( 152 )

References | Related Articles | Metrics

This paper proposes a method to inpaint damaged old photos based on generative adversarial networks. The generator is based on the U-Net network and uses partial convolution instead of all convolutional layers. It only operates on effective pixels, which not only avoids the color discrepancy and blurriness caused by standard convolution, but also can repair irregular damaged area. Considering the dependence on long-distance feature information, the contextual attention model is added in the decoding stage of generation network to maintain semantic coherence. In addition to the basic GAN loss, the loss function of the generator also adds perceptual loss, style loss and reconstruction loss to enhance network stability. Experiments are conducted on the CelebA-HQ dataset and real damaged old photos. The experimental results show that the method is not limited by the damage and can achieve a good restoration effect on the damaged old photos.

Design and Implementation of Micro-site for Digital Rural by Assist of UAV Aerial Photography

[HS(]LONG Jing-jing, HUANG Lei, GOU Wen-yu, MENG Wen-hao, LI Ming-feng, CHEN Fu-chang, HUANG Cheng, HE Zhi-min, LI Yan, LIN Hui-chuan

2021, 0(04): 48-52.

Asbtract ( 150 )

PDF (7543KB) ( 91 )

References | Related Articles | Metrics

Constructing digital rural is an effective measure to bridge the gap between urban and rural areas. It is a necessary measure to realize rural informatization and modernization. It is an important part of realizing rural revitalization strategy and building a well-off society in an all-round way. In this paper, by using drone aerial photography to obtain orthophoto images with longitude and latitude, and then stitching these images by Photoscan and ArcMap, a real image map with centimeter resolution has been produced. And on this basis, a digital rural service platform integrating geographic information and digital resources has been constructed. Taking Zhangtang village in Dongshan county, Fujian province as an example, we explore the construction scheme of such a digital rural integrated service platform. Moreover, a “Digital Zhangtang” micro-site, which can be accessed through mobile scanning programs such as WeChat, Alipay, browser website, has been developed.

Improved YOLOv3 Vehicle Detection Algorithm Embedded in Dilated Convolution Module

HU Chang-ran, FAN Yan-guo, YU Ding-feng

2021, 0(04): 53-60.

Asbtract ( 282 )

PDF (7207KB) ( 188 )

References | Related Articles | Metrics

Vehicle detection on image or video data is one of the hotspots in the field of computer vision, and it is also an important part of intelligent transportation systems. In view of the complex and changeable vehicle detection scenes and the existing vehicle detection algorithms can not meet the requirements of high precision and high real-time at the same time, this paper proposes an improved YOLOv3 vehicle detection algorithm and builds its own vehicle detection data set. First, we embed the dilated convolution module in the original and feature extraction network Darknet-53 to reduce the loss of target information and enhance the receptive field. Secondly, in the NMS (non-maximum suppression) module, in order to reduce the missed detection, this article discusses the traditional NMS and makes improvements. If the IoU of the prediction frame is greater than the set threshold, it will be attenuated in a certain way. The improved method shows better performance than other algorithms on the KITTI standard data set, and the verification accuracy can reach 96% in the self-built data set, and the detection speed is 25.9 frames/s.

Flower Recognition Based on ResNet and Attention Mechanism

ZHANG Meng-yu

2021, 0(04): 61-67.

Asbtract ( 398 )

PDF (1698KB) ( 201 )

References | Related Articles | Metrics

Flower recognition has important application value in life, and the traditional flower recognition methods have some problems, such as low recognition accuracy and weak generalization ability. To solve these problems, this paper proposes a ResNet34 network model with attention mechanism. After the first convolutional layer and each residual block of ResNet34, channel attention mechanism and spatial attention mechanism are added, and the transfer learning is used for training network model. Experiment shows that ResNet34 has a higher recognition accuracy rate than AlexNet, VGG-16 and GoogLeNet on the flower data set. The ResNet34 model with attention mechanism and transfer learning has 6.1 percentage points higher recognition accuracy than the original model, and 1.1 percentage points higher recognition accuracy than the original model with transfer learning only. Compared with traditional deep learning models, the model proposed in this paper significantly improves the recognition accuracy.

Performance Analysis of mmWave Communication via Aerial Large Intelligent Surfaces

CHENG Yin-xuan, ZHOU Si-yuan, TAN Guo-ping, ZHANG Zhi, ZHAN Jia-li

2021, 0(04): 68-73.

Asbtract ( 134 )

PDF (1190KB) ( 95 )

References | Related Articles | Metrics

As a new type of reflective materials, large intelligent surface (LIS) will play a vital role in the next generation networks. By deploying LISs mounted on unmanned aerial vehicle (UAV), UAV-LIS can be used to further enhance the network coverage of LIS. In this paper, we propose a communication network model based on UAV-LIS and analyze the dual-hop transmission of millimeter wave downlink communication. Specifically, we use Poisson point process to depict the locations of users served by UAV-LIS which cruises around a base station. Then we derive the probability density function of the distance from the ground users to UAV-LIS. The analytical expression of system coverage probability can be obtained and verified by the simulation results. It turns out that the combination of LIS and UAV can greatly improve the coverage performance of millimeter wave communication system in urban areas. The developed framework is insightful for determining the network configuration by which the optimum system performance can be achieved.

WSN Coverage Optimization Based on GABC of Feature Point Set

HOU Yi-fei, YANG Yong

2021, 0(04): 74-78.

Asbtract ( 158 )

PDF (1020KB) ( 95 )

References | Related Articles | Metrics

In order to solve the problem that the artificial bee colony algorithm uses grid points to calculate the network coverage, which will lead to a large amount of calculation and is easy to fall into the local optimal solution, a global optimal solution based on feature points set is proposed to optimize wireless sensor networks. Firstly, the target area is divided into a limited number of feature points, and the coverage of the sensor is transformed into the coverage calculation of several feature points, which reduces the calculation of coverage rate and describes the coverage of the whole network. Then, on the basis of feature point set, the global optimal solution artificial bee colony algorithm is successfully applied in the field of network coverage, and the performance of standard artificial bee colony algorithm and artificial bee colony algorithm based on global optimal solution in network coverage is compared. Simulation experiment results show that after optimizing node coverage based on the global optimal solution artificial bee colony algorithm, the coverage rate is effectively improved and it is not easy to fall into the local optimal solution.

Measurement Node Selecting Method Based on Improved Greedy Algorithm

WU Shang, SHENG Yi-qiang, DENG Hao-jiang,

2021, 0(04): 79-84.

Asbtract ( 150 )

PDF (936KB) ( 84 )

References | Related Articles | Metrics

In the future application scenarios, there is a need of deterministic delay guarantee for the name resolution system. How to effectively select measurement nodes and provide support for exact delay name resolution is the problem this article focuses on. This paper maps the network measurement scheduling deployment problem to the minimum vertex cover problem and proposes an improved greedy algorithm which mainly optimizes the greedy algorithm iteration cycle and improves the sorting algorithm in certain scenarios. The experimental results show that the improved greedy algorithm is more than 900% faster than traditional greedy algorithm.

An ICN Cooperative Cache Strategy Based on Node Centrality Approximation Algorithm

LUO Lan-hua, YUAN Shu-dan, HE Qiao-ping

2021, 0(04): 85-90.

Asbtract ( 117 )

PDF (1196KB) ( 83 )

References | Related Articles | Metrics

In order to reduce the cache redundancy and average access cost of information centric networking (ICN), a collaborative cache strategy based on node centrality metric approximation algorithm (CMAA) is proposed. Considering the impact of the workload of accurately calculating the shortest path on the cache performance, CMAA uses the approximate estimation of the shortest path to improve the calculation efficiency of node centrality, and takes the weighted fusion value of the approximate measurement of node centrality, node heat and cache utilization rate as the cache impact factors to calculate the cache priority of each node in the interest packet forwarding path. The simulation results of CMAA under various experimental conditions show that compared with LCE (leave copy everywhere) and CLFM (cache “less for more”), CMAA can effectively improve the cache hit rate, reduce the average access cost and improve the performance of the cache system with little change in the average cache request delay.

A New Efficient and Lightweight Convolutional Neural Network Model

ZHANG Jian-jian

2021, 0(04): 98-103.

Asbtract ( 158 )

PDF (2592KB) ( 101 )

References | Related Articles | Metrics

Aiming at the problem that the current food recognition system has a large number of network model parameters and a large model, this paper proposes a 23-layer network model with only 204k parameters. The basic building blocks (7×7, 5×5, 3×3) are used to generate feature maps, and two pooling layers of different receptive fields are used to fuse the feature map of the convolutional layer, and a 1×1 convolution kernel is used for nonlinear combination. Then it is connected to the spatial pyramid pooling layer, and finally is classified in the softmax classifier. Experiments on public data sets show that, compared with ResNet50 and GoogLeNet, the network model in this paper reduces model parameters by 99.14% and 96.63% respectively without reducing classification performance.

Dynamic Load Balancing Algorithm for Heterogeneous Clusters Based on Internet of Things Technology#br#

LI Juan

2021, 0(04): 104-108.

Asbtract ( 145 )

PDF (1968KB) ( 86 )

References | Related Articles | Metrics

In order to improve the dynamic load balance of heterogeneous clusters, the sensor technology in the Internet of things technology is introduced to construct a channel transmission model of heterogeneous clusters, and then the output channel is configured by dynamic weighting method, and the channel characteristics are decomposed. On the basis of establishing the channel fuzzy reorganization structure model, the noise interference suppression method is used to suppress the multipath interference of heterogeneous cluster communication channel. Combining with baud interval balanced sampling method to control the balance of channel output, the dynamic load balancing of heterogeneous cluster is realized through ambiguity balanced configuration and space balanced scheduling process. The simulation results show that the load scheduling of heterogeneous clusters is well balanced, the bit error rate of output signals is reduced, and the output balance and adaptive control ability of heterogeneous clusters are improved.

Design and Implementation of Security Experiment of Internet of Things Based on Smart Phone and Smart Home Appliances for New Engineering Disciplines#br#

ZHANG Wen-bo, HU Xi-ming, LI Peng, MA Miao, WANG Tao,

2021, 0(04): 109-116.

Asbtract ( 153 )

PDF (2952KB) ( 99 )

References | Related Articles | Metrics

Internet of things (IoT) accelerates to form a strategic pillar industry from a pioneering industry. The collaborative development of IoT production and education urgently needs to carry out experimental technology innovation for new engineering fields. Through the investigation and analysis of the experimental teaching literatures of IoT in China from 2005 to 2020, it is found that the experimental mode is evolving from principle verification to independent exploration, the experimental environment is upgrading from large-scale site to portable experience type, and the experimental teaching reform is oriented towards smart home appliances. On this basis, this paper proposes the IoT security experiment technology based on “mobile phone+smart home appliances”, gives the experiment process, transforms personal mobile into “experimental weapon” through the key technology of “Linux+Tools+Python”, completing the experiments of mDNS amplification and reflection attack and replay attack of mobile phones against smart home appliances and their attack defense. This technology effectively guarantees online experimental courses teaching during the epidemic period, and provides a new technical way for new engineering to actively develop the IoT industry and deepen the reform of experimental teaching in universities.

Black Box Adversarial Attack Algorithm Based on Deep Reinforcement Learning

LI Meng, HAN Li-xin

2021, 0(04): 117-121.

Asbtract ( 337 )

PDF (973KB) ( 133 )

References | Related Articles | Metrics

Aiming at the problem of black box adversarial attack in the field of image recognition, a black box adversarial attack algorithm is proposed based on the DDQN framework and Dueling network structure in reinforcement learning. The agent generates an adversarial sample by imitating human adjustment of the image, interacts with the attacked model to obtain misclassification results, and calculates the structural similarity of the clean sample and the adversarial sample to generate a reward. During the attack, only the label output information of the attacked model was obtained. The experimental results show that the success rate of attacking the four deep neural network models trained on the CIFAR10 and CIFAR100 datasets exceeds 90%. The quality of the generated adversarial samples is similar to the white box attack algorithm FGSM and the success rate is more advantageous.

Spam Recognition Method Based on BiGRU-Attention-CNN Model

ZHAO Yu-xuan, HU Huai-xiang

2021, 0(04): 122-126.

Asbtract ( 178 )

PDF (800KB) ( 93 )

References | Related Articles | Metrics

E-mail is an important communication tool, but the problem of spam has been affecting peoples daily work and life. Continuously improving spam detection technology and increasing the speed and accuracy of spam detection has important research and practical significance. Bi-directional gated recurrent unit (BiGRU) and convolutional neural network (CNN) are widely used in the field of text classification. The combination of them could give full play to BiGRU context dependency extraction capabilities and CNN feature extraction capabilities. But for the problem of spam recognition, it is also necessary to consider some specific words in the email. So this article proposes a spam recognition method based on the BiGRU-Attention-CNN model to improve the accuracy of spam detection. The model first converts the email text into feature vectors and performs BiGRU serialization learning, and then introduces the attention mechanism (Attention) to give greater weight to specific words. After the attention layer is input to the CNN model, through convolution, pooling, and full connection, the classification result is finally obtained. The model is tested on the Trec06c mail data set and compared with other models, better results are achieved. The final accuracy of the model is 91.62%.

Table of Content