毕业论文外文翻译--并条机自调匀整利用人工神经网络确定在自调匀整作用点(外文原文+中文翻译)

Textile Research Journal Article Use of Artificial Neural Networks for Determining the LevelingAction Point at the Auto-leveling Draw FrameAssad Farooq1and Chokri CherifInstitute of Textile and Clothing Technology, TechnischeUniversität Dresden. Dresden, GermanyAbstractArtificial neural networks with their ability of learning from data have been successfully applied in the textile industry. The leveling action point is one of the important auto-leveling parameters of the drawing frame and strongly influences the quality of the manufactured yarn. This paper reports a method of predicting the leveling action point using artificial neural networks. Various leveling action point affecting variables were selected as inputs for training the artificial neural networks with the aim to optimize the auto-leveling by limiting the leveling action point search range. The Levenberg Marquardt algorithm is incorporated into the back-propagation to accelerate the training and Bayesian regularization is applied to improve the generalization of the networks. The results obtained are quite promising.Key words:artificial neural networks, auto-lev-eling, draw frame, leveling action point。

The evenness of the yarn plays an increasingly significant role in the textile industry, while the sliver evenness is one of the critical factors when producing quality yarn. The sliver evenness is also the major criteria for the assessment of the operation of the draw frame. In principle, there are two approaches to reduce the sliver irregularities. One is to study the drafting mechanism and recognize the causes for irregularities, so that means may be found to reduce them. The other more valuable approach is to use auto-levelers [1], since in most cases the doubling is inadequate to correct the variations in sliver. The control of sliver irregularities can lower the dependence on card sliver uniformity, ambient conditions, and frameparameters.At the auto-leveler draw frame (RSB-D40) the thickness variations in the fed sliver are continually monitored by a mechanical device (a tongue-groove roll) and subsequently converted into electrical signals. The measured values are transmitted to an electronic memory with a variable, the time delayed response. The time delay allows the draft between the mid-roll and the delivery roll of the draw frame to adjust exactly at that moment when the defective sliver piece, which had been measured by a pair of scanning rollers, finds itself at a point of draft. At this point, a servo motor operates depending upon the amount of variation detected in the sliver piece. The distance that separates the scanning rollers pair and the point of draft is called the zero point of regulation or the leveling action point (LAP) as shown in Figure 1. This leads to the calculated correction on the corresponding defective material [2,3]. In auto-leveling draw frames, especially in the case of a change of fiber material, or batches the machine settings and process controlling parameters must be optimized. The LAP is the most important auto-leveling parameter which is influenced by various parameters such as feeding speed, material, break draft gauge, main draft gauge, feeding tension, break draft, and setting of the sliver guiding rollers etc.Use of Artificial Neural Networks for Determining the Leveling Action Point A. Farooq and C. CherifFigure 1 Schematic diagram of an auto-leveler drawing frame.Previously, the sliver samples had to be produced with different settings, taken tothe laboratory, and examined on the evenness tester until the optimum LAP was found (manual search). Auto-leveler draw frame RSB-D40 implements an automatic search function for the optimum determination of the LAP. During this function, the sliver is automatically scanned by adjusting the different LAPs temporarily and the resulted values are recorded. During this process, the quality parameters are constantly monitored and an algorithm automatically calculates the optimum LAP by selecting the point with the minimum sliver CV%. At present a search range of 120 mm is scanned, i.e. 21 points are examined using 100 m of sliver in each case; therefore 2100 m of sliver is necessary to carry out the search function. This is a very time-consuming method accompanied by the material and production losses, and hence directly affecting the cost parameters. In this work, we have tried to find out the possibility of predicting the LAP, using artificial neural net-works, to limit the automatic search span and to reduce the above-mentioned disadvantages.Artificial Neural NetworksThe motivation of using artificial neural networks lies in their flexibility and power of information processing that conventional computing methods do not have. The neural network system can solve a problem “by experience and learning” the input–output patterns provided by the user. In the field of textiles, artificial neural networks (mostly using back-propagation) have been extensively studied during the last two decades [4–6]. In the field of spinning previous research has concentrated on predicting the yarn properties and the spinning process performance using the fiber properties or a combination of fiber properties and machine settings as the input of neural networks [7–12].Back-propagation is a supervised learning technique most frequently used for artificial neural network training. The back-propagation algorithm is based on the Widrow-Hoff delta learning rule in which the weight adjustment is carried out through the mean square error of the output response to the sample input [13]. The set of these sample patterns is repeatedly presented to the network until the error value is minimized. The back-propagation algorithm uses the steepest descent method, which is essentially a first-order method to determine a suitable direction of gradient movement.OverfittingThe goal of neural network training is to produce a network which produces small errors on the training set, and which also responds properly to novel inputs. When a network performs as well on novel inputs as on training set inputs, the network is said to be well generalized. The generalization capacity of the network is largely governed by the network architecture (number of hidden neurons) and this plays a vital role during the training. A network which is not complex enough to learn all the information in the data is said to be underfitted, while a network that is too complex to fit the “noise” in the data leads to overfitting. “Noise” means variation in the target values that are unpredictable from the inputs of a specific network. All standard neural network architectures such as the fully connected multi-layer perceptron are prone to overfitting. Moreover, it is very difficult to acquire the noise free data from the spinning industry due to the dependence of end products on the inherent material variations and environmental conditions, etc. Early stopping is the most commonly used technique to tackle this problem. This involves the division of training data into three sets, i.e. a training set, a validation set and a test set, with the drawback that a large part of the data (validation set) can never be the part of the training.RegularizationThe other solution of overfitting is regularization, which is the method of improving the generalization by constraining the size of the network weights. Mackay[14] discussed a practical Bayesian framework for back-propagation networks, which consistently produced networks with good generalization.The initial objective of the training process is to mini-mize the sum of square errors:∑=-=ni i i D a t E 12)( （1）Where i t are the targets and i a are the neural network responses to the respective targets. Typically, training aims to reduce the sum of squared errors F=Ed.However, regularization adds an additional term, the objective function,W D E E F αβ+= （2） In equation (2), w E is the sum of squares of the network weights, and α and βare objective function parameters. The relative size of the objective function parameters dictates the emphasis for training. If α<< β, then the training algorithm will drive the errors smaller. If α>>β, training will emphasize weight size reduction at the expense of network errors, thus producing a smoother network response [15].The Bayesian School of statistics is based on a different view of what it means to learn from data, in which probability is used to represent the uncertainty about the relationship being learned. Before seeing any data, the prior opinions about what the true relationship might be can be expressed in a probability distribution over the network weights that define this relationship. After the program conceives the data, the revised opinions are captured by a posterior distribution over network weights. Network weights that seemed plausible before, but which do not match the data very well, will now be seen as being much less likely, whilethe probability for values of the weights that do fit the data well will have increased[16].In the Bayesian framework the weights of the network are considered random variables. After the data is taken, the posterior probability function for the weights can be updated according to Bayes’ rule:(3) In equation (3), D represents the data set, M is the particular neural network model used, and w is the vector of network weights.),/(M w P α is the prior probability, which represents our knowledge of the weights before any data is collected. ),,/(M w D P βis the likelihood function, which is the probability of data occurring, given the weights w.),,/(M w P βαis a normalization factor, which guarantees that the total probability is 1 [15].In this study, we employed the MATLAB Neural Net-works Toolbox function “trainbr” which is an incorporation of the Levenberg–Marqaurdt algorithm and the Bayesian regularization theorem (or Bayesian learning) into back-propagation to train the neural network to reduce the computational overhead of the approximation of the Hessian matrix and to produce good generalization capabilities. This algorithm provides a measure of the network parameters (weights and biases) being effectively used by the network. The effective number of parameters should remain the same, irrespective of the total number of parameters in the network. This eliminates the guesswork required in determining the optimum network size.ExperimentalThe experimental data was obtained from Rieter, Ingolstadt,the manufacturer of draw frame RSB-D40 [17]. For these experiments, the material selection and experimental design was based on the frequency of particular material use in the spinning industry. For example, Carded Cotton is the most frequently used material, so it was used as a standard and the experiments were performed on carded cotton with all possible settings, which was not the case with other materials. Also, owing to the fact that all the materials could not be processed with same roller pressure and draft settings, different spin plans were designed. The materials with their processing plans are given in Table 1.The standard procedure of acclimatization was applied to all the materials and the standard procedure for auto leveling settings (sliver linear density, LAP, leveling intensity) was adopted. A comparison of manual and automatic searches was performed and the better CV% results were achieved by the automatic search function from RSBD-40.Therefore the LAP searches were accomplished by the Rieter QualityMonitor (RQM). An abstract depiction of the experimental model is shown in Figure 2.Use of Artificial Neural Networks for Determining the Leveling Action Point A. Farooq and C. CherifFigure 2 Abstract neural network model.Here the point to be considered is that there is no possibility in the machine to adjust the major LAP influencing parameter, i.e. feeding speed. So feeding speed was considered to be related to delivery speed and number of doublings according to equation (4). The delivery speed was varied between 300 and 1100 m/min and the doublings were 5 to 7, to achieve the different values of the feeding speed:Feeding Speed=Delivered Count× Delivery speed/（Doublings ×Feed Count）（4）Training and Testing SetsFor training the neural network, the experimental data was divided into three phases. The first phase included the experimental data for the initial compilation of the data and subsequent analysis. The prior knowledge regarding the parameters influencing LAP, i.e. feeding speed, delivery speed, break draft, gauges of break and main draft, and the settings of the sliver guide, was used to select the data. So the first phase contained the experiments in which the standard settings were taken as a foundation and then one LAP influencing parameter was changed in each experiment.In the second phase, the experiments were selected in which more than one influencing parameter was changed and the network was allowed to learn the complex interactions. This selection was made on the basis of ascertained influencingparameters, with the aim to increase or decrease the LAP length. The third phase involved the experiments conducted on the pilot scale machine. These pilot scale experiments were carried out by the machine manufacturer to get the response for different settings. So these results were selected to assess the performance of the neural networks.Pre-processing of dataNormalizing the input and target variables tends to make the training process better behaved by improving the numerical condition of the problem. Also it can make training faster and reduce the chances of getting stuck in local minima. Hence, for the neural network training, because of the large spans of the network-input data, the inputs and targets were scaled for better performance. At first the inputs and targets were normalized between the interval [–1, 1], which did not show any promising results. Afterwards the data was normalized between the interval [0, 1] and the networks were trained with success.Neural Network TrainingWe trained five different neural networks to predict the LAP by increasing the number of training data sets gradually. The data sets were divided into training and test sets as shown in Table 2. The training was performed with training sets and test sets were reserved to judge the prediction performance of the neural network in the form of error. Figure 3 depicts the training performance of the neural network NN 5.Figure 3 Training performance of NN 5.Results and DiscussionAs already mentioned in Table 2, different combinations of the data sets were used to train five networks keeping a close look on their test performance, i.e. performance on the unseen data. As Bayesian regularization is said to eliminate the guess work for the appropriate number of hidden layers and hidden neurons, the numbers of hidden neurons, ranging from 14 to 22, were selected in two layers to train the networks. The basic idea behind the selection of network topology is that the network should be complex enough to learn all the relationships in the data as the possibility of overfitting was tackled with regularization. The behavior during training and performance on the test sets can be seen from Figures 4–8.Figure 4 Testing performance of NN 1.Use of Artificial Neural Networks for Determining the Leveling Action Point A. Farooq and C. CherifFigure 5 Testing performance ofNN 2.Figure 6 Testing performance of NN 3.Figure 7 Testing performance of NN 4.Figure 8 Testing performance of NN 5.Mean absolute error was calculated to compare the predicted and actual values, both during training and testing, and the results are presented in Figure 9. The overall results can be explained by considering the presence or absence of input–output interactions in the training data and the increase in the prediction performance with the increase of training data sets. Figure 9 clarifies the increase in testing error, even with an increase in the number of training, data sets, when training and testing was performed on data sets from different phases (NN2 & NN4, see Table 2). However, the error showed a downwards trend when part of the phase data was used to train the network and the remainder was used for testing, as in NN 3 and NN 5 in comparison with NN 2 and NN 4, respectively. The presence of different input–output interactions in different phases explains this trend. The exceptional behavior of NN 1 with respect to the above-mentioned fact is attributed to the relatively small number of data sets in phase 1.Figure 9 Error comparison between training and test sets.In order to assess the goodness of fit of NN 5, a 10-fold cross-validation was preformed, i.e. using 90% of the data for training and 10% for testing, repeating the training 10 times and testing the network each time on 10% of the unseen data. An average 2R= 0.9622 was reported. The same procedure was adopted for the 80% training and 20% test sets and the calculated value of 2R was 0.9470. This decrease in the performance is due to the 80% data sets available for training. However, these values confirm a very good fit of the NN 5 model.ConclusionThe artificial neural network model was developed and the networks were trained at the Institute of Textile and Clothing Technology, Technische Universitat Dresden. The use of Bayesian regularization to reduce the testing error for the practical applications has shown quite promising results. From the testing performance of NN 5 as shown in Table 3, a maximum deviation of about 2 mm is observable, which falls well within the 3 mm negligible range for determination of the LAP. This concludes that neural networks can be applied in the future for quick computation of the LAP, with the advantages of fast adjustment and saving of material and time. The accuracy in computation can lead to better sliver CV% and better yarn quality.Table 3 Test performance of NN 5.中文翻译纺织研究期刊并条机自调匀整利用人工神经网络确定在自调匀整作用点摘要用他们的人工神经网络从数据中学习的能力已经成功的应用在纺织行业，在并条机自调匀整参数中匀整作用点是一个重要的参数，并且强烈的影响到纱线的质量，本文主要是讲述的是用人工神经网络来检测匀整作用点，各种各样的匀整作用点会影响导致不同的变量，我们把这些变量输入到人工神经网络进行测试，目的是为了通过限制匀整作用点改变范围来优化并条机的自调匀整，Levenberg-Marquardt的算法是纳入反向传播加速系和贝叶斯正则化方法,改进了网络的泛化。