Feature extraction and recognition technique for human pulse signals

In Chinese traditional medicine, the pulse signal plays an important role for diagnosis. To investigate the effective technique for identifying heroin druggers through the pulse signal, the feature extraction algorithm based on the biocepstrum and the third-order cepstrum entropy of the pulse signal is studied. On the basis of concise and rigorous discussion for the algorithm, the biocepstrum-based diagonal slice components are estimated for human pulse signals of 20 heroin druggers and 20 healthy normal subjects. The magnitude of the sample value of diagonal slice at m=n=1 and the third-order cepstrum entropy of magnitudes of sample values of the diagonal slice within a particular region are used as two feature parameters for every human pulse signal to form a feature vector. A classifier based on the criterion of squared Mahalanobis distance is successfully designed. Applying the designed classifier to 40 feature vectors, the correct identification rate reaches 87.5%. The research result shows that the method of the feature extraction and classifier design presented in is valuable for identifying the human pulse signals of heroin druggers.     

Application of the Wavelet Transform to the Pulse Signal Analysis of Druggers

According to the randomness of human pulse signals,the multiresolution analysis of the wavelet transform is used to analyze such signals.Its purpose is to extract the abnormal information from the pulse signals of heroin druggers and to obtain the primary judgment criterion which can be used to identify druggers from healthy persons.The scale spectrum based on the wavelet transform of pulse signals carries the important characteristic information of the health situation of human body.The pulse signals of 15 heroin druggers and 15 healthy persons are analyzed and the scale spectrum and the total signal energy of every signal are extracted.It is found that the ratio between the sum(i.e.,scale-wavelet energy) of the scale spectrum in a specific scale-time region and the total signal energy for heroin druggers is generally higher than that of healthy persons.Using the percentage of the ratio between the scale-wavelet energy in the specific scale-time region and the total signal energy as characteristic parameter,a critical parameter is determined that is used to classify heroin druggers and healthy persons.Thus,all of the 15 healthy persons are identified correctly from 30 subjects.Only two heroin persons are misjudged.The experiment results of classification show that the method presented is feasible and effective for detecting the pulse abnormalities of heroin druggers.     

基于布谷鸟搜索神经网络的微波加热温度预测模型

微波加热是一种与被加热物直接相互作用的选择性加热方式,具有清洁、节能、减排等特点。针对工业物料作为微波加热负载时,其温度非线性变化的特点,以微波工业加热过程中的多维、海量参数为研究对象,基于泛函接神经网络模型提取样本数据的深度特征,提出了一种基于布谷鸟搜索算法,优化BP神经网络的网络参数,建立了以"数据驱动"为手段微波加热工业物料温度模型。仿真实验结果证明了所提出模型的准确性、实时性。     

BP神经网络在2005年太湖蓝藻水华预警中的应用

蓝藻水华是目前中国乃至世界面临的重大环境问题之一。为了有效地减少及预防蓝藻水华带来的影响,收集苏州市吴县1986—2007年的气象资料,运用主成分分析法,分析了太湖蓝藻暴发前一个月的主要限制因子及其相互关系。气温、气压、相对湿度、降水是影响叶绿素a浓度的主要限制因子。结合2005年1—10月太湖各区域蓝藻叶绿素a浓度的含量,利用Matlab R2010a软件,建立了基于BP神经网络的蓝藻水华预警模型,可为采取相应措施和控制蓝藻水华提供科学依据。     

遗传算法优化人工神经网络模型在日光温室湿度预报中的应用

提出一种基于遗传算法优化BP神经网络的方法预测日光温室湿度环境因子。实测日光温室内影响空气湿度的环境因子组成数据样本作为神经网络的输入,采用基于实数编码的遗传算法替代随机设定神经网络的初始权阈值,然后通过改进的BP算法在由遗传算法确定的搜索空间中对网络进行精确训练。模型预报值和实测值基于1：1线的决定系数R2和预测平均相对误差MSE分别为0.9857和3．1％。结果表明,遗传算法优化BP神经网络预报模型收敛速度快、预测精度高。可为日光温室的湿度环境调控制提供理论依据和决策支持。     

Pulse Signal Analysis of Druggers with the Wavelet and Neural Network

The most significant difference between the human pulse signals collected from heroin druggers and healthy persons is at their amplitude waveforms as time functions. That is, the amplitude values and change rates of two types of signals, within a particular time range, appear different features. However, the partial components of the scaling and wavelet coefficients of the pulse signals obtained by using wavelet transform can reveal such key features. The pulse signals of 15 heroin druggers and 15 healthy persons are analyzed through using the muhiresolution analysis of wavelet transform. By using db2 orthogonal wavelet, every pulse signal is decomposed into three levels and the absolute values of the sixth component of scaling coefficients and the second component of the wavelet coefficients in the third level are combined to form a feature vector. A probabilistic neural network with good detection performance is successfully designed for automatically detecting 30 feature vectors. During the network design, 20 feature vectors are used as training samples. The remained 10 feature vectors are used as testing samples. Based on these steps, 15 heroin druggers and 15 healthy persons are all correctly identified. In other words, the detection rate arrives at 100%. druggers.     

Application of Intelligent Neural Network on Prediction of Time Series Signal

The research of neural network has been maturated both in theory and practical application since 1980's, and also been employed into the prediction and analysis of nonlinear time series signal in the field of signal process system. Concerning with the problem of time series signal prediction based on traditional neural network, such as black box, poor accuracy, and facing the shortage of post knowledge, this paper presents a different neural network prediction model from the traditional ones, based on intelligent neural cell model and employing the iterative prediction method. Through the example on stock price prediction, the prediction accuracy and practical value are proved.     

基于数值预报产品的地表温度预报方法研究

为了开展地表温度预报业务,提高逐日地表温度预报准确率,利用2007—2012年的ECMWF和T213数值预报产品资料及抚顺市的逐日地表温度资料,采用逐步回归分析方法和BP神经网络模型分别构建抚顺市地表温度预报模型,并对模型的精度进行检验。结果表明,地表温度与ECMWF的高度场、海平面气压场、温度场和T213的散度场、高度场、海平面气压场、地面气压场、海平面K指数、水汽通量、相对湿度、温度场、地面气温和场涡度场均呈显著相关。对预报模型进行精度检验显示,地表平均温度和地表最低温度的预报效果较好,≤3℃预报准确率均达到79%以上。2种模型对比显示,BP神经网络预报模型总体上优于逐步回归预报模型;逐步回归预报模型较BP神经网络预报模型稳定。     

Thermal process model identification based on radial basisfunction neural networks

In order to accurately reflect the dynamic behavior and realize the whole optimal control of the thermal process, a novel modeling method of the RBF NN (Radial Basis Function Neural Networks) model is proposed to build nonlinear model. This method is based on entropy clustering and competitive learning algorithm, combined with nonlinear autoregressive moving average (NARMA) model to identify the RBF NN stucture, and the power vector is gotten by the least square algorithm. Two simulation experiments show that the proposed method of the identification based on NARMA model and RBF NN can accurately describe the non linearity of the process and has less hidden nodes.     

Fast algorithm for apparent resistivity calculation based on back propagation neural networks

To avoid the complex numerical calculation for the electromagnetic field and determine underground abnormality, a neural network based method is proposed. In consideration of turn off transmitter current, the effect of a linear ramp turn off current on transmitter is corrected. The characteristics of transient expression and the traditional calculation algorithm for apparent resistivity are analyzed, and a predigest structure of network is obtained based on the kernel expression. The three layer back propagation(BP) neural network is trained by using sample data in homogeneous half space, and its number in hidden layer was determined. The method proposed is compared with two traditional calculation methods with simulation experiments. The result demonstrates that BP neural network has a high speed of processing data and is useful in explanation of the transient electromagnetic method.     

A method for evolving the architecture and weights of neural networks via gene expression programming

An algorithm for automatic designation of the architecture and the weights of neural networks using gene expression programming (GEP) was presented. The fundamental ideas and procedures of the algorithm were discussed. The algorithm was improved to solve the problems of prematurity and lower variance rate. An application for neural networks designation was given. The experimental results indicate that the proposed GEP approach may evolve the architecture of neural network, and can obtain the weights more precisely. Compared to other conventional evolutional algorithms, GEP shows faster convergence.     

Study on Intelligent Lighting Design Method

In this paper,the artificial neural network theory is applied to indoor lighting design,the lighting design is combined with lighting calculation, which is fairly quick and reasonable.Furthermore this lighting design method is characterized by self_fitting and self_studying.     

BP神经网络在烟蚜发生程度预测中的应用

为实现对烟田烟蚜发生程度的预测预报,以12年的历史资料为基础数据,采用BP神经网络方法建立了烟蚜发生程度的预测模型。该模型对待测样本的预测准确度为99.43%,回测准确度为87.36%。所建立的预测模型可提前1个多月对烟蚜发生程度进行预测,为中期预测模型,其预测结果可为烟田蚜虫综合治理提供依据。     

Fetal electrocardiogram extraction based on radial basis function neural networks

A novel method for extracting fetal electrocardiogram (FECG) from the abdominal composite signal of a pregnant woman is proposed. The maternal component in the abdominal electrocardiogram (ECG) signal is a nonlinearly transformed version of the mother's ECG (MECG). This nonlinear relationship was identified using radial basis function (RBF) neural networks. The FECG is extracted by subtracting the nonlinearly transformed version of the MECG from the abdominal ECG signal. The baseline shift and noise in the FECG are suppressed by wavelet packet denoising technique. Experimental results obtained from the actual ECG signals demonstrate the effectiveness of the proposed method in extracting FECG even when it is totally embedded within the maternal(QRS) complex.     

BP neural network and genetic algorithm for thefilling properties optimization of crankshaft

The structure of wedge flash is proposed to improve the filling properties for deep cavity structure of crankshaft die with dimensional splitting mold. BP genetic algorithm is applied to optimize the structure parameters of wedge flash based on Matlab. Samples which are selected by orthogonal test are analyzed via FEM, and the minimum unfilled distance obtained are employed to conduct the BP neural network training. Then the optimum parameters with minimum unfilled distance are gained from genetic algorithm. Error between the parameters predicted and the results get from simulations is less than 5%. The productive practice indicates that the cavity is fully filled and the material utilization ratio increases from 75.7% to 81.4%, which confirms the correctness of optimization of wedge flash structure.     

Fault diagnosis of rotating machinery based on self-learning fuzzy spiking neural network

For fuzziness classific boundry of fault diagnosis of rotating machinery and traditional neural network algorithms difficulted to solve contradiction between application problems example scale and netwok scale,a methord of self-learning fuzzy spiking neural network is put forward. The methord overcomes unavailability of cluster analysis on classific boundry of fault diagnosis of rotating machinery by species encoding of pulse sequence and unsupervised learning. The method shows that it effectively solves boundary fuzziness problem on fault diagnosis of rotating machinery,and greatly improves efficiency of fault diagnosis.     

Cost estimate of power line projects based on grey relational analysis and neural networks

To accurately estimate the cost of a power line project,a method based on grey relational analysis (GRA) and neural networks (NN) is presented and studied. Grey relational analysis technologies are used to analyze the features of the transmission line project and ten main features which affect the project cost most are selected. Then, the main features are used as input neural cell of neural networks, and a model of GRA-ANN is built. To verify the method, the cost data of a 110 kV power construction project are used to train and test the model. Results show the model’s maximum relative error of static investment is 3.72% and the minimum is 1.85%, and its accuracy is high. The LM-BP algorithm and the traditional BP algorithm are used respectively to train the GRA-ANN network, and results show the error declining rate of LM-BP algorithm is faster and the overall error is lower.