Study On The Prediction And Deduction Model Used In The Virtual Manufacturing Of The Worsted Textiles

With the promising future of the modern science and technology, the manufacturing technology with the automated and mechanized modes, which was ever satisfied when the category of the products with the huge yields was seldom varied, trends to the flexibility, systematization, and intelligence because the sorts of the products is frequently altered while their demand are often in a few quantity. The virtual manufacturing is one of the milestones in the modern advanced processing technologies, which has been widely researched and developed in the fields of the manufacturing industries such as machine, automobile, aviation etc. However, in textile industry of China, the latest science achievements are not widely used and researched. Especially, the advantages on raw materials and labor cost disappeared after China accessed into the WTO more than 8 years. At the same time, the variety of the species, small amount and quick consignment have become the fashion in the textiles trades. Therefore, it provides a fortunate chance for Chinese textile industries to greatly improve the textiles manufacturing through the modern advanced processing methods, information technology and innovation ability, which is also an important industry project eagerly required to be resolved now.Consequently, the virtual manufacturing technology (VMT) of worsted textiles has been researched and discussed based on the practical database from one representative wool mill. Using the rough set theory (RS) to analyze those data and get rules, based on case based reasoning (CBR) the case and rule based reasoning (CBR&RBR) had been discussed. Genetic algorithms (GA) was used to optimize the weightiness and threshold of artificial neural network (ANN) prediction and deduction models. It is possible for the manufacturers to realize the prediction and control of the quality, the adjustment of parameters, the design of the new products and the decision-making of the produce plans through VMT, which can improve the research level of the information and intelligence of China textile industry. Meanwhile, the basic theory and applied technology, which are utilized to optimize the VTM through the different methods, are valid and effective in this thesis. According to our work and efforts, four exciting achievements are shown as follows.1. Mixed reasoning of case-based & rule-based and rough predictionIn order to deal with the problems of the empiricism and variety of the new textile design, the most similar case used should be firstly retrieved according the main characteristic parameters of the textiles based on case-based reasoning algorithm. According to retrieving, the most similar cases are as references for a new product design process. In view of the characteristics of rough set theory, rough set data analysis (RSDA) has been primarily studied in order to obtain knowledge rules. Based on the simplified attributes and extractive rules, another case was established using rule-based reasoning. RSDA was combined with CBR to build the case library directly. This allowed unimportant parameters to be removed and the case library to be simplified; in turn allowing easier, more efficient searching of attributes from the simplified case library. The union of Rule Based Reasoning (RBR) and CBR means that complicated calculations around similar cases and the associated error can be avoided. So it is possible to realizing the rough prediction by quickly searching case library and extracting the most similar case from the RBR& CBR hybrid case library.2. Combination of hybrid reasoning technology and ANN model to implement quality controlIn view of the aforementioned hybrid technology of case-based and rule-based reasoning, the most similar cases which was retrieved from the case library based on the RBR&CBR, was used to achieve the rough prediction of the processing procedures. At the same time, using the optimized ANN forecasting model, the process parameters of the most similar case was tested. If the production quality met the requirements, then directly using the technology to manufacture; otherwise according to the forecast suggest, the sensitive parameters cab be adjust that lead to optimizing the process programs, which was in order to realizing rapid design and quality control.Moreover, according to the reverse flow of product design, the technological parameters can be determined. In this paper, the reverses deduce models of main process established by ANN were references for mixture of production resources. For the deduce model of spinning parameters, the average prediction accuracy of the yarn draft ratio, ring traveler and spindle speeds are higher than 97%; For the deduce model of top parameters, the average prediction accuracy of the fiber mean diameter, coefficicient of variablity of diameter, fiber mean length, coefficicient of variablity of length and short fiber content are higher than 95%; For the deduce model of after finishing parameters, the average prediction accuracy of the wash variable, cook variable and steam variable are 90.33%,95.90% and 80.35%.3. Using principal component and factor analysis to select the network input parametersBased on results of the principal component eigenvalues and the explaining variance, the input parameters of artificial neural network prediction model (ANN) were selected. There is a rule that the most important parameter group or parameter was firstly input, less import parameter was input later. So all selected main factors were input into ANN model. Compared with other screening method:subjective experience (SE) and multiple stepwise regression (MSR), the accuracy and stability of prediction model are best. When using PCA&FA method, input parameters of ANN model reduced; the number of the hidden layer nodes reduced correspondingly; the topological structure of network was simplified. Thus the learning rate of network was improved, the network prediction performance increased to a certain extent.PCA&FA, not only can optimize the network input parameters, but also discriminate the sensitive parameters. For pre-spinning processes, the more sensitive parameter group were fiber characteristics (fiber mean diameter (x3), coefficient of variability of diameter (x4), fiber mean length (x5), coefficient of variability of length (x6)) and the drafting state of tops(oil content of tops (x1), total drawing ratio (x12)); for spinning processes, the more sensitive parameter group were the spinning technological parameter(yarn draft ratio (x14), ring traveler (x15) and nominal warp count (x18)) and nominal warp twist (x17); for weaving processes, the most sensitive parameter group was the weaving technological parameter(heddle frame height (x27), opening height (x28), back-beam height (x29) and loom speed (x20)); for after-finishing processes, the more sensitive parameter group were fabric structure parameter(float length (x23) and real weft density (x33)) and fabric twist(nominal warp twist (x17) and nominal filling twist (a1)).4. Using genetic algorithm to optimize network initial weights and thresholdsIn view of the characteristics of GA and BP algorithm, this paper proposed a hybrid training program:the initial selected weights and thresholds randomly of BP were replaced a better search space determined by GA optimization. When the optimized weight and threshold matrixes were input into the network, the network was trained secondly by the Levenberg-Marquardt (LMBP) algorithm. Among the GA optimization, the chromosomes encoding and fitness function were designed; the three genetic operations:selection, crossover and mutation were introduced in detail. After the optimization of network weights and thresholds by GA, network rate of convergence was increased and the error sum squares further decreased. Using the trained model to predict the validation data, the average predict accuracy has increased to some extent.For roving process, the average accuracy of roving CV (R1) and weight (R2) increased by 0.03% and 0.43% respectively, which is due to prediction accuracy without GA optimization is already very high (more than 97%). For spinning process, the average accuracy of eight variables improved on different degrees:from 0.11% for the yarn CV(Y1) (without GA optimization the prediction accuracy was over 97%) to 5.15% for yarn thin places per kilometers (Y3), which the previous accuracy 88.18% increased to 92.72%. For weaving process, the average accuracy of weaving efficiency (W1) and woven defects (W2) increased 1.96% and 3.63% respectively. For after finishing process, the average accuracy of eight variables improved on different degrees:from 0.07% for the seam shrinkage of weft (F6) (without GA optimization the prediction accuracy was 90.21%) to 13.45% for the steam shrinkage in width (F4), which the previous accuracy 72.63% increased to 82.40%.In summary, this thesis proposed a hybrid reasoning techniques and realized rough prediction of the main worsted processing based on the combination of case-based reasoning (CBR) and rough set theory (RS). Further, the virtual manufacturing models were established of the overall process of worsted textiles based on combined principal component analysis and factor analysis (PCA&FA) with artificial neural network (ANN). And these prediction models were further optimized by genetic algorithms (GA). At the same time as the combination the rough prediction of the optimized virtual manufacturing model, the integrated virtual manfacturing system had been established. So the rapid design of process of products and quality control were feasible.