On-line measurement and characterization of yarn and fabric qualities using a wavelet-stochastic hybrid method

This thesis is concerned with new approaches to measurement, monitoring and characterization of textile manufacturing processes and products by using a combination of stochastic and wavelet models and analysis methods.; The results are aimed at controlling yarn defects and visual qualities of the resulting fabrics. The stochastic models facilitate detection and identification of spinning faults, while use of wavelet analysis provides a compact representation of signal features without significant loss of information. A scheme has been developed, tested, and applied to extract, retain, and synthesize only the essential information required for characterizing the salient qualities of yarns and fabrics, without having to process and store vast amounts of data acquired on-line. More specifically, the system developed is capable of (i) performing data reduction by on-line screening of unnecessary information, (ii) localizing quality features, in time domain, of abrupt changes including undesirable faults signal features, (iii) generating a large amount of simulated data from the compressed data by combining the wavelet packet analysis and stochastic simulations, and (iv) producing a sufficient area of fabric images for prediction of visual qualities. The hybrid system based on the stochastic and wavelet analyses was applied to a large amount of data from an actual yarn manufacturing process, and the performance was verified by applying the system developed. Ten yarn packages were measured on a Shlafhorst{dollar}spcircler{dollar} open-end spinning machine at normal production conditions, providing data for a total of length of 250,000m, or 25,000m for each package. The massive amount of data could be reduced to a small fraction consisting of (i) a set of statistical parameters for normal sub-blocks, (ii) sets of statistical parameters on abnormal sub-blocks for characterizing yarn and fabric defects and for providing their time-domain addresses, and (iii) wavelet coefficients for abnormal sub-blocks. The reduction rate was approximately 99.9999% in the example at which 100 spindles could be monitored for 10 hours by storing only a small fraction of the data. For maximizing the reduction rate, the wavelet coefficients were further discarded retaining only the minimum and maximum values at each scale. It was also shown that the algorithm developed for data reduction can be used to generate a variety of virtual yarns and fabric images that are almost identical to that from the original signals while retaining all important time-specific information of the faults.