Applied Mechanics and Materials Vols. 548-549

Abstract: Because fatigue monitoring based on the image of the non-contact measurement is single and low accuracy, a novelty driver fatigue monitoring system based on multivariate hierarchical Bayesian network is proposed. The system mainly includes four modules following: face region detecting, eyelid closure judging, head region positioning, and fatigue analyzing. The eye region is positioned precisely by the method of gray projection function, the binary image which contains the whole eye information using self-adapting threshold method is obtained, and then driver fatigue monitoring system based on hierarchical Bayesian network is used to evaluate the fatigue level of the driver. The experimental results show that the fatigue monitoring accuracy is up to 90% in specific conditions, it’s effective to improve the detection accuracy compared to the other method.

Abstract: With the vigorous development of research-oriented learning in the field of education, more and more educators and institutions begin research and develop the research-oriented learning platform based on web, in order to improve the students' autonomous learning ability. But the platform itself or the information of users given exist many problems, cannot achieve the intended purpose. This paper introduces data mining technology to the research-oriented learning platform building.

Abstract: Feature selection is an important process to choose a subset of features relevant to a particular application in document classification. Those terms which occur unevenly in various categories have strong distinguishable information as to categorization. Firstly, based on the categorical document frequency probability (CTFP), a CTFP_VM feature selection algorithm was designed for feature selection. Secondly, a maximum term frequency conditional distribution factor was proposed to improve the CTFP_VM criterion further. We perform the document categorization experiments on SVM classifiers with the well-known Reuters-21578 and 20news-18828 corpuses as unbalanced and balanced corpus respectively. Experiments compare the novel methods with other conventional feature selection algorithms and the proposed method achieves the excellent feature set for document categorization.

Abstract: Facial expression recognition is a key ingredient to either emotion analysis or pattern recognition, which is also an important component in human-machine interaction. In facial expression analysis, one of the well-known methods to obtain the texture of expressions is local binary patterns (LBP) which compares pixels in local region and encodes the comparison result in forms of histogram. However, we argue that the textures of expressions are not accurate and still contain some irrelevant information, especially in the region between eyes and mouth. In this paper, we propose a compound method to recognize expressions by applying local binary patterns to global and local images processed by bidirectional principal component analysis (BDPCA) reconstruction and morphologic preprocess, respectively. It proves that our method can be applied for recognizing expressions by using texture features of global principal component and local boundary and achieves a considerable high accuracy.

Abstract: Mixed expression is more in line with people’s daily perfromance than basic expression.This paper proposed a facial expression recognition method that recognizes and analyze mixed expressions. In this method, Gabor phase and local binary patterns were combined into GPLBP model to obtain the expression features and the model contained good robustness of light. Compressed sense and subjection degree function were adopted to identify the ingredients of main basic expressions in the mixed expression the ratio of each kind of basic expression. Experimental results of comparison respectively to Gabor-SVM and AAM-SVM verified that the proposed method could not only identify and analyze the mixed expression effectively but also recognize the basic expression precisely.

Abstract: This paper proposes a non-uniform rational B-spline (NURBS) curve extraction algorithm from 2D laser sensor data and 3D simulated data. In robot localization and mapping application, the raw sensor data cannot be stored due to its need of large storage space. However, only a small number of control points of NURBS curve are needed to be stored to recover the geometrical feature of raw data. To comprise the number of control points and accuracy of the extracted curve, global approximation method is adopted to minimize the error between the raw data and the extracted curve. In extraction process, all the weights are set as one firstly. After find the control points, a weight calculation method is developed to update the weight values. The NURBS curve with new weight has smaller error than with original weights. Finally, NURBS curve extraction results from real 2D laser sensor data and 3D simulated data are shown to check the feasibility of proposed algorithms.

Abstract: Symbolic aggregate approximation is data dispersed dimension reduction method. Time series are equal and unequal length after dimensional reduction through difference methods. For calculating similarity of unequal-length symbolization time series, a new method is proposed. First, key point method is used to process time series for getting important information. Then, the new method of this paper is done to let them equal in local. Finally, SAX is utilized to symbolic time series and then calculate similarity of them. The experimental results show that this method is simple and effective and it extends the field of calculating the similarity of unequal length time series.

Abstract: Load model has a great impact on the digital simulation result. In this paper, the measurement-based method is applied to model the load. If all the measured data are used for modeling respectively, the workload would be increased greatly. But if only one model is generated with the multi-curve fitting parameter identification method, the accuracy of modeling would be reduced greatly. The clustering analysis theory supplies an effective way to solve the problem above. There are some methods for clustering introduced in this paper. But a suitable method needs be studied firstly. The case study is presented to compare these methods. According to simulation result, it is concluded that the Kmeans method is best, while the usually adopted central clustering is actually not suitable for the load time-variation research.

Abstract: Self-sensing control of permanent magnet synchronous motor has been a hot research topic of motor control technology, which has the advantages of low cost and high reliability. The use of a high-frequency carrier signal superimposed on the fundamental excitation of PMSM has established itself in recent years as a viable means of eliminating position sensors in ac drives for applications. In the carrier signal injection methods, the rotor position can be estimated by using the carrier signal current resulting from the interaction between the carrier signal voltage and the spatial saliency. The existing sampling error and random noise of high frequency current signal were uncertain to the rotor position self-sensing, that may affect the rotor position self-sensing precision. This paper describes a new wavelet de-noising method of the high-frequency current signal, improving the de-noising results significantly.

Abstract: This paper uses OpenMP parallel standardsin massive data Delaunay triangulation.Based on the Fork/Join parallel mode in OpenMP, a quad-tree is designed to divide and map the point cloud data.And the hierarchical implementation of triangulation and merging operations are based on it. A WFM-JLP algorithm is used to schedule triangulationand merging operations in order to achieve better load balancing. Experiment show the proposed method can greatly reduce the real-time memory and computing time.