Search

The metal defect detection and identification technology have been successfully applied in actual industrial production. But there are still some problems in the steel strip defect detection task, such as the difficulty in data collection and the poor identification effect caused by the small defect target. Aiming at these problems, this paper presents an algorithm based on convolutional neural network for the detection of steel strip small target defects. First, the original image is clipped by the way that moving the slider. At the same time, the feature pyramid structure is increased to conduct feature fusion for the output of the feature extraction part and Group Normalization is used to accelerate the convergence of the model, so that the model can better adapt to the target task, effectively solve the problem of small defect object information loss on the surface of the target steel strip, and enhance the generalization of the model. In order to evaluate and test the validity of the model, a steel strip defect dataset (BS5-DET) is constructed. The experimental results show that the mAP of the model reaches 54.50% on the defect data set of the BS5-DET steel belt, which has a certain application value.

In this paper, a practical test system of relay protection device is designed that considered from practical characteristics and demands. The test system used the existing software technology and short circuit calculation principle. The practice has proved that the system can complete common functions such as whole group test, open-in test, open-out test protection logic test and so on. The system can meet all the test needs of the relay protection products and improve the efficiency and quality of the testing. It’s helpful to study the comprehensive impact of system failure on all the devices in the station.

Text in natural scene images are often faced with low-resolution problem, which brings significant difficulties to many text-related tasks such as text detection and recognition. In this paper, we propose a novel text-attentional Conditional Generative Adversarial Network (cGAN) model for text image super-resolution (SR). The model enhances the original cGAN by introducing effective channel and spatial attention mechanisms based on the proposed Residual Dense Channel Attention Block and text/non-text segmentation information, which focus the model on the text regions instead of the background of the image to learn more effective representations of text and achieve better text super-resolution result. The proposed model achieves state-of-the-art performances on public text image super-resolution dataset.

Critical nodes of software systems are the nodes with the ability to transmit fault rapidly in the whole system. As a result, they have a significant impact on software reliability. To evaluate and identify the critical nodes, previous research focused on the static structure of software network which is inadequate since dynamic information affects the evaluation result as well. We propose a comprehensive method combining both static and dynamic information to evaluate critical nodes in our work. Static information consists of parameters measuring software structure. The importance of one node differs in different structures. In our evaluation model, we choose betweenness centrality to measure static structure because it can describe the importance of one node in static topological structure in complex network theory. For dynamic information, we focus on the execution frequency in specific software operation profile and user related operation. Combining the static and dynamic evaluating results, the comprehensive model is obtained. Using this evaluation model, we can obtain the critical nodes for one software system corresponding to different service states which indicate some sets of operating states. Protection strategy can be applied to these nodes pertinently to achieve efficient system reliability improvement.

The nanocrystalline Si quantum dot (nc-Si QD) stacking MOS structure is fabricated by plasma-enhanced chemical vapor depositon (PECVD). Negative differential resistance (NDR) characteristics are investigated in the nc-Si quantum dot floating gate MOS structure. Clear multi NDR peaks for electrons and holes are observed in the I-V curves and calculations indicate that these NDR characteristics should be associated to Coulomb blockade effect and quantum confinement effect of the nc-Si quantum dots.

Based on the single electron tunneling and storage in an uniform nc-Si single layer within a-SiN x (or a-SiO x ) / nc-Si/a-SiN x (or a-SiO x ) structures, the nc-Si floating gate nonvolatile memories (NVM) have been fabricated. The performance of the program and erase speed as well as the retention time have been improved by using the nitrogen plasma treatments. In order to increase the storage density, the double level charge storage has been realized in the stacked nc-Si layer based MIS structures. Considering the scaling limit of charge storage in nc-Si NVM, the Si high-rich SiO x based electrically switchable resistance random access memory (RRAM) has been investigated. The mechanism of electrically switchable resistance in Pt/SiO x /Pt structure has been also discussed.

Centered around the model of ESDDM, a software evolution process model based on evolution behavior interface, and the knowledge of component classification and component retrieval evaluation, this paper furthered previous work by proposing a component reusability evaluation method based on fuzzy mathematics. The method provides an effective way of evaluating the reusable ability of a certain component during the process of software evolution.

Software evolution requirement, which is the cross point of software evolution and software requirement, is the source power of software evolution. In order to improve the efficiency and quality of software evolution, this paper modeled the software evolution requirement process. Based on the study of software evolution, software requirement and feedback mechanism, this paper proposed two different types of feedbacks in software evolution requirement process, which perfected the change control process of software evolution requirement. After bringing in colored dual-transitions Petri net (CDPN), process model of software evolution requirement (SERPM) based on feedback was formalized. SERPM can effectively manage the changing process of software evolution requirement and thus software can evolve with high efficiency and quality.

To improve the identification precision of flow-rate characteristics of pneumatic solenoid valves, a new algorithm called Cbm algorithm is proposed. The error formation mechanism of current algorithm defined by the ISO6358/DIS-3 is analyzed. The Cbm algorithm has reduced the errors of pressure derivative by changing the calculation procedure of multivariable optimization. Five pneumatic solenoid valves are tested to compare the identification precision of different algorithms. The comparison results show that the Cbm algorithm improves the identification precision of flow-rate characteristics. The identification errors of sonic conductance of the Cbm algorithm are less than 4%, and the identification errors of critical pressure ratio of the Cbm algorithm are less than 10%.

Capacitance-voltage (C-V) and frequency dependent conductance-voltage (G-V) measurements have been carried out to investigate the charging and discharging effect induced by interface states and nanocrystalline Si (nc-Si) in floating gate MOS structures. Distinct conductance peaks are observed in the G-V curves for the floating gate with and without nc-Si dots. Based on the calculation of interface states density and charge density, the role of nc-Si and interface states is analyzed in detail.

The nc-Si nonvolatile memory devices with high performance have been fabricated by using general CMOS techniques. High resolution transmission electronic microscope (HRTEM) shows that the average size of nc-Si is 8 nm and its density is 3×1011/cm2. The performance of programming/ erasing and retention time is mainly depending on the quality and thickness of tunnel layer and control layer. The results show the relation between the performance and novel design of tunnel and control layers.

Size-controlled nanocrystals (Si NCs) in floating gate MOS structure have been fabricated by thermal annealing of Si-rich SiNx layers with high ratio of Si/N. High resolution transmission electronic microscopy (HRTEM) reveals the size of Si NCs can be controlled by varying the thickness of Si-rich SiNx layer. Based on the analysis of XPS and Raman measurement, the relation between the size of Si NCs and the thickness of Si-rich SiN x is discussed.

The costs of pre-building graphic in the construction project management is a very important job, based on neural network theory and the characteristics of the costs of building graphic, raised GA neural network to solve the cost of building graphic problem, applied BP network model and used GA to optimize the weight based on BP algorithm. The calculation examples showed that the accuracy of the cost estimation met the requirements basically.

In this paper, course timetabling problem was referred as the research object, the objective function converted by many constraints and soft constraint condition was analyzed. Established the mathematical model, proposed discrete particle swarm algorithm and simulated annealing algorithm that combined to settle course timetabling problem, which ensured that the particles escaped from the local minimum, guaranteed that the curriculum be more close to the ideal, the experimental results showed that the better results can be obtained by the method.

In this paper, the backpropagation algorithm-the multilayer feedforward network structure is described . The proposed approach is experimented, tested to classify corporate financial performance using data with financial ratios Financial modeling for classification algorithms based on neural network is established. The study found that Levenberg Marque training error is smallest among 4 learning algorithms and its performance is better.

The different types of classification algorithms for credit scoring systems are used to evaluate credit risk. Neural network-based systems that allow the system, through an analysis of historical data, to determine the relationship between account characteristics and the probability of default. The backpropagation algorithm -the multilayer feedforward network structure is described based on data with nine financial ratios from 81 firms listed in china. A simulation on network is made. Neural network model for classification algorithms are established. By varying network parameters we demonstrate that Levenberg Marque training error is smallest among 4 learning algorithms and its performance is better. Increasing the number of hidden layer can result in minor improvement.

Neural Network-based systems that allow the system, through an analysis of historical data ,to determine the relationship between account characteristics and the probability of default. The backpropagation algorithm ? the multilayer feedforward network structure is described .based on data with nine fiancial ratios from 81 firms listed in china, A simulation on network is made . Neural Network model for classification algorithms are established . By varying Network Parameters we demonstrate that LevenbergMarque training error is smallest among 4 learning algorithms and its performance is better. Increasing the number of hidden layer can result in minor improvement.

There exist many customers in credit market that needs to be classified into distinct groups. K-Means Algorithm are presented, which based on the historical financial ratios, utilizing the cluster analysis technology to analyze the listed enterprises in Zhejiang province. Some indicators related to financial attributes are analyzed, and nine finance indicators are chosen. According to better valuation on the companies listed, we apply to "try and error" and choose 4 as the number of clustering. 81 samples are divided into two groups: one training group with 60 firms and other testing group with 21 samples. Testing results shows that the model trained can be available for clustering companies listed in Zhejiang province.