Research On Intelligent Mining And Quality Control Techniques And Its Application In Textile Production Process

Within textile fields, application of digital technology has made rapiddevelopment, from general business management to computer control in productionand computer aide design in textile. One of the main characters in modem textiletechnologies is the intelligent control and smart production way. The increasingquality demands from the customers make clear the need to research novel ways ofquality control. The goal is to monitor in-process data as a means of improvingquality and reducing production cost. So this thesis makes an attempt to solve suchproblems as follows:ⅰ. Intelligent mining architecture and system modeling toward textile productionⅱ. Support vector machines based textile processing quality predictionⅲ.Knowledge representation and knowledge discovery from complex textile processⅳ. Intelligent methods of textile quality controlWith the integration of artificial intelligence, data mining and automationtechniques, this thesis explored innovative theories and methods on rational textileprocess, and developed an intelligent mining software system adapt to textileproduction. All in all, the main important research is as follows:There has been proposed a multiple-agent based intelligent mining architecture(MIMA) for textile process decision and quality control, including data-preprocessingagent, data mining agent, knowledge evaluation agent, knowledge service agent,central control agent and human-machine agent. The advantages of the systemarchitecture lie in robust reasoning, various services and the hybrid mining algorithms.Under the MIMA, the system utilities, intelligence, and capability of self-learning areimproved, and the system complexity is reduced to some extent. The data miningplatform (DWTP), that is data warehouse for textile production, has been designedand built to effectively organize, extract, store and process industry data.Novel support vector machines (SVM) are employed for developing textile process prediction models, which are so-called support vector machines toward textileproduction (SVMT~2-P). Theν-support vector regression machines with radial basiskernel function have been designed to build up the prediction models. Performance ofthe SVMT~2-P is estimated by the k-fold cross-validation method. Optimization ofSVM parameters such as the sparsity parameterν, kernel widthσ, and penaltyterm C is performed with genetic algorithms. Under the SVM's approach, three kindsof different textile predictive models have been developed respectively, namely,cotton yarn, air-jet texture yarn and worsted yarn quality model. The predictivepowers of the SVMT~2P are estimated by comparison with the artificial neuralnetworks (ANN) models. The experimental results indicate that in the small data setsand real-life production, the SVMT~2P is capable of improving predictive accuracy upto 4～17%, and more suitable for noisy and dynamic spinning process.Under the MIMA, the hybrid representations of textile process knowledge havebeen studied in this thesis, including object-oriented case, rough sets (RS), supportvector machines or artificial neural networks, and so on. The theoreticalunderstanding of ANN models which are based on minimization of the generalizationerror help to increase the degree of confidence of their use. New rough sets basedtextile knowledge acquirement method has been investigated, and its computing flowalso demonstrated. In addition, case based knowledge reuse techniques have beendiscussed, such as for case indexing, case adaptation, and case storage.An intelligent Control Model toward Textile Quality (ICMT~2Q) has beendesigned to optimize textile process and raw material. The ICMT~2Q has coupleclosed-loop systems which can create so-called feedback control on the textile quality.Discrete decision tables on textile process are built by the means of theself-organizing maps (SOM) in the neural network field. As a result, the quality ruleshided in process data are induced, which can be used to diagnose quality problemswithin real production. The reasoning mechanism of the ICMT~2-Q has been studied inthis thesis. The case investigation reveals that the ICMT~2Q is sufficiently transparentfor spinners to understand how steady process quality is obtained.On the basis of theory study, supported by the national technology innovation fund, a web based intelligent process planning & virtual manufacturing system(WIPVM 1.0) has been developed. The software system is made up of threesub-systems, i.e., textile process planning & knowledge management, dataacquirement system from textile production, textile quality prediction & virtualmanufacturing system. By the engineering application of software system, it isobserved that the new theories and methods for textile process optimization andquality improvement are promising.The main contributions of this thesis include:1. Novel support vector machines toward textile production (SVMT~2P) have beenproposed for quality prediction and process optimization. Under the SVM's approach,three kinds of different textile predictive models have been developed respectively,namely, cotton yarn, air-jet texture yarn and worsted yarn quality model.2. A multiple-agent based intelligent mining architecture (MIMA) has been built upfor textile process decision and quality control. Under the MIMA, the hybridrepresentations of textile process knowledge have been proposed. New knowledgeacquirement methods such as rough set and case based reasoning have beeninvestigated, and their computing flows also demonstrated.3. A new control model (ICMT~2Q) has been proposed to optimize textile process andraw material. The ICMT~2Q has couple closed-loop control systems, which canprovide an alternative approach for textile quality intelligent control.