学位论文 > 优秀研究生学位论文题录展示

Modeling of Drafting Force and Sliver Unevenness

作　者: NIBIKORA ILDEPHONSE
导　师: Wang Jun
学　校: 东华大学
专　业: Textile Engineering
关键词: Drafting Force Sliver Unevenness Linear multiple regression MLP neural network RBF neural networks GMDH polynomial neural networks GRNN Cascade correlation neural networks Genetic Programming Smoothing spline
分类号: TS114.2
类　型: 博士论文
年　份: 2011年
下　载: 10次
引　用: 0次
阅　读: 论文下载

内容摘要

Draw frame contributes less than 5% to the production costs of yarn; however its influence on yarn evenness is very high. Further, if the draw frame is not properly adjusted, there will also be effects on yarn strength and elongation. The increasing productivity and keep the sliver regularity in the draw frame depend on the many factors like the quality of raw material, level of modernization and condition of machinery, effective humidification control, use of optimum process parameters, efficient house-keeping and careful materials handling. The draw frame reduces the linear density of sliver by passing the sliver through successive pair of rollers rotating at higher speeds. In order to establish the drafting process conditions, it is required to describe the relationship between process variables and material parameter with dynamic states and online device for the drafting process. Many factors affect the size and variation of drafting force and sliver quality. Some can be changed in steady-state and others in the dynamic state. The objective of this work is to analyze the effect of drafting process parameters on the drafting force and sliver quality which can be changed under dynamic conditions. Since drafting force is measured on-line, and sliver unevenness is measured offline, once we can prove their relationship, it is possible to get real-time information about sliver quality (on-line inspection).Firstly, a new type of draw frame (DHU-A301) with two weighting sensors and a frequency-change speed regulator are used to set the machine and measure the drafting force. The study used two factors (draft distribution and speed of the front and back roller) to analyze the size and variation of drafting force and sliver unevenness. The relationship between rotational speed of the back and front roller and frequency rate was estimated. Linear regression, the most widely used method, and which is well understood, is used to analyze experiment results. Main draft and break draft increase the drafting force and sliver unevenness and on the other hand, front roller speed reduces the drafting force. According to the sensitivity analysis of linear multiple regression model, break draft is the most important factors for drafting force and sliver unevenness. To evaluate this model we used the root mean squared error (RMSE), mean absolute error (MAE) and the mean absolute percentage error (MAPE). A satisfactory result for drafting force was achieved; however, linear regression didn’t prove satisfactory prediction accuracy for sliver unevenness. Therefore, there is a need to use soft computing methods like artificial neural networks, which can give good result for nonlinear systems.Secondly, different approaches of neural networks were designed with the same variables. All approaches have their advantages and disadvantages and it is therefore necessary to know their different characteristics in order to choose the best possible approach to model a system. The goal was to minimize the root mean square error. There are so many neural networks which can be trained, starting by Multilayer perceptron neural network, because it is the most popular used of many different research applications. Multilayer perception network provides satisfactory drafting force prediction accuracy from the drafting process parameters, but it didn’t provide satisfactory prediction accuracy for sliver unevenness. Like linear multiple regression, break draft is the most important factors for drafting force and sliver unevenness. The next was Radial Basis Function neural networks, because it can be trained much faster than the MLP neural network, and it can also give better performance with minimum number of training data set. The results show that Radial Basis Function neural networks have improved the prediction accuracy compare to MLP networks (R-square equal to 0.90) for drafting force but it didn’t achieve the performance accuracy for sliver unevenness. Also like previews methods, break draft is the most important factor for drafting force and sliver unevenness. Furthermore, Group Method of Data Handling (GMDH) polynomial neural networks were trained because the network structure is flexible and amenable to topological search, and also it offers adaptive network representations. This method didn’t achieve a good result for drafting force but it has improved the prediction accuracv for sliver unevenness and also, the most important factor is the front roller sneed. Another method is the. General Regression Neural Network (GRNN). because it is the fast learning and converges to the optimal regression surface as the number of samples become very large. The other very big advantage is that they are very flexible and new information’s can be added immediately with almost no retraining. This method can achieve the best result for both drafting force and sliver unevenness. The other method trained was Cascade correlation neural networks; because its training time is very fast and the numbers of layer and neurons to be used in the network are self organizing, no need to adjust the parameters. The comparison of those methods shows that General Regression Neural network is the best for prediction improvement of drafting force and for sliver unevenness and the Cascade correlation neural networks can be selected for predicting sliver unevenness. The analysis of importance proves that the break draft is the most important variable except for Cascade correlation neural networks where the speed of front roller is important.Thirdly, Artificial Neural Networks were used in the preview chapter because it is the most widely used method in textile modeling especially in the nonlinear system. This chapter use Genetic programming to train the same data with the same variables because it has many advantages over ANN like, it can generates equations or formulas relating to inputs and output. The analysis and prediction of signal output for drafting force measurement on the draw frame can be in the future the main parameter to control online sliver quality. Genetic programming shows a very good agreement between experimental and predicted values (R-square=0.9169). The result obtained from genetic programming is a good tool for developing a new drafting force measurement. Fourthly, the work aim is to study the relationship between drafting force and sliver unevenness. Two approaches were used, simple linear regression and smoothing spline method to analyze the relationship between drafting force and sliver unevenness. The work used the principle that there is linear relationship between drafting force and stress-strain of the strain gages in the sensor, and self-developed set of drafting force online inspection device. Using this device can accurately measure the size of the drafting force, then the data obtained can be used to investigate the relationship between drafting force and sliver unevenness. The system can provide real-time information about the sliver quality, but there is a problem of noise data which need to be solved. The result shows that there is a smoothing spline relationship between drafting force and sliver unevenness.Finally, application of neural networks (General regression neural networks) to the visual inspection of drafting force. In the absence of human operators, this function has to be performed with intelligent decision-making systems that are able to interpret incoming sensor information and decide on the appropriate control action. Intelligent decision-making systems are expected to replace the knowledge, experience, and the combined sensory and pattern recognition abilities of human operators. Successful implementation of these different tasks depends on two factors; first, the quality of information obtained from the monitoring sensor, and second, the techniques used to process this information in order to make decisions. The main objective is to provide a solution to overcome this problem through the development of a decision support system with an expert system component, able to minimize the subjectivity of the human expert in monitoring and behavioral prediction of drafting force measurement. General Regression Neural Networks provide satisfactory performance for forecasting future values of drafting force from previews values.

全文目录

ACKNOWLEDGEMENTS  5-6
Abstract  6-9
Contents  9-13
Tables  13-16
Fiugres  16-19
1 Introduction  19-23
  1.1 Motivation  19-20
  1.2 Objective  20-21
  1.3 Outline  21-22
  References  22-23
2 Literature review  23-52
  2.1 Introduction  23
  2.2 Operating principle of draw frame  23-32
    2.2.1 Operating devices  25
    2.2.2. Requirement of drafting arrangement  25-26
    2.2.3 Element of drafting arrangement  26-28
    2.2.4 Drafting arrangement  28-29
    2.2.5 Influences on the draft  29
    2.2.6 Monitoring and Autolevelling  29-32
  2.3 Overview of research on the draw frame  32-48
    2.3.1 Ideal drafting  32-34
    2.3.2 The theory of floating fibres  34-35
    2.3.3 Division of the total draft into sectional drafts  35-39
    2.3.4 The theory of drafting force  39-40
    2.3.5 Research on the Drafting force  40-42
    2.3.6 Drafting process parameters  42-43
    2.3.7 Sliver unevenness  43-47
    2.3.8 Minimum irregularity and index of irregularity  47-48
  References  48-52
3 Effect of drafting process parameters on the drafting force and sliverunevenness  52-70
  3.1 Introduction  52
  3.2 Methods  52-58
  3.3 Linear regression  58-59
  3.4 Experimental procedure  59-63
  3.5 Results and Discussions  63-68
  3.6 Conclusion  68
  References  68-70
4 Performance improvement prediction accuracy of drafting force andsliver unevenness using neural networks models  70-114
  4.1 Multilayer Perceptron Neural network  70-80
    4.1.1 The Multilayer Perceptron Neural Network Model  71-72
    4.1.2 Training Multilayer Perceptron Networks  72-74
    4.1.3 Experiment and results  74-80
  4.2 Radial Basis Function neural networks  80-87
    4.2.1 RBF Network Architecture  80-81
    4.2.2 Training RBF Networks  81-82
    4.2.3 Experiment and results  82-87
  4.3 GMDH polynomial networks  87-92
    4.3.1 Structure of GMDH network  87-89
    4.3.2 GMDH Training Algorithm  89
    4.3.3 Experiment and results  89-92
  4.4 General Regression Neural networks  92-100
    4.4.1 Architecture of a PNN/GRNN Network  93-94
    4.4.2 GRNN training Algorithm  94-96
    4.4.3 Experiment and results  96-100
  4.5 Cascade Correlation Neural Networks  100-108
    4.5.1 Cascade Correlation Network Architecture  100-101
    4.5.2 Training Algorithm for Cascade Correlation Networks  101-103
    4.5.3 Experiment and results  103-108
  4.6 Comparison of all the predictions models  108-110
  4.7 Conclusion  110-111
  References  111-114
5 Using Genetic Programming to forecast drafting force  114-127
  5.1 Introduction  114
  5.2 Advantages of GP  114-115
  5.3 Basic considerations  115-116
  5.4 Solving problems with GP  116-123
  5.5 Implementations  123
  5.6 Experimental results and discussions  123-125
  5.7 Conclusion  125
  References  125-127
6 Relationship between drafting force and sliver unevenness  127-135
  6.1 Introduction  127
  6.2 Methods  127-131
  6.3 Data analysis  131-133
  6.4 Conclusion  133-134
  References  134-135
7 Application of neural networks to visual inspection of drafting force  135-141
  7.1 Introduction  135
  7.2 Time series  135-136
  7.3 GRNN  136-137
  7.4 Experiments results and discussions  137-139
  7.5 Conclusion  139
  References  139-141
8 Conclusions and Recommendations  141-144
  8.1 Conclusion  141-143
  8.2 Recommendations  143-144
Appendix  144-162
List of Publications  162

Modeling of Drafting Force and Sliver Unevenness

内容摘要

全文目录

相似论文