Abstract:

Objective The yarn breakage in ring spinning directly affects the production efficiency and product quality of the yarns. At present, the commonly used automatic yarn break detection mainly employs a single spindle detection method with photoelectricity or magnetoelectricity as the core, both requiring costly modification of the spinning machine and is difficult to implement. Therefore, this paper proposes a break detection method based on machine vision, which provides a new direction for achieving low-cost and high-precision break detection.

Method An online detection method for yarn breaks based on visual feature reinforcement extraction was proposed. For the problem of difficult yarn feature extraction due to yarn trailing, a neighborhood gradient reinforcement operator was designed for yarn clustering to achieve yarn feature reinforcement. To deal with the problem that yarn targets are small and easily disturbed by environment, an Otsu small target segmentation threshold search method inspired by valley bottom was designed to achieve the segmentation of yarn and background, to extract adaptively yarn features from the yarn image after feature reinforcement, and to enable broken end detection by Euler distance discriminant method.

Results Inspection devices were installed at a textile factory in Wuxi to collect data, with 1 000 captured images selected for analyze. The factory mainly produces pure cotton high count yarns, with 400 spindles per vehicle. To verify the superiority of the weak feature enhancement proposed in this research, the proposed algorithm was compared with Retinex, homomorphic filtering, and histogram averaging for enhancement experiments. The algorithm developed in this research enhanced the yarn features and effectively suppressed the background areas based on the gradient step characteristics of the yarn (Fig. 8 (e)). Choices of different weights for neighborhood gradient reinforcement operators had different effects on the separation of yarn features and background features. In order to select the optimal weight value, this research adopts the method of controlling variables and conducts experimental comparison on the center weight value between 2 and 3. It was found that the optimal selection range of weight values was between 2.4 and 2.7. Distinct image and dim image enhancement renderings under different weight show clear and blurred images, respectively.

To verify the effectiveness of the algorithm in extracting yarn features, Otsu, Otsu with added pixel proportion weight, and Otsu threshold segmentation method inspired by valley points were applied to the images. Using this algorithm, it was demonstrated that the proposed algorithm effectively extracted yarn features and a small number of background features at the same pixel level. It was also be demonstrated that the proposed algorithm was able to search effectively for valley thresholds and achieve yarn feature extraction.

To verify the effectiveness of the feature extraction algorithm, the processing effects and detection results of the Hough transform line extraction method, LSD line detection algorithm, Sobel, Robert and Otsu combined algorithm, Linknet algorithm, and this algorithm were compared(Fig. 13). The results obtained by this algorithm were best fitted to the original image (Fig. 13 (g)).

The experimental results showed that the detection rate of this method reached 97.3% and the processing efficiency reached 59.76 ms per frame, which was carried out by collecting 1 000 spinning pictures of a spinning factory. It was evident that this method was able to effectively detect yarn breakage in real-time.

Conclusion The algorithm reported in this paper decomposes the acquisition of yarn features into two parts: reinforcement and extraction, solving the problems of yarn feature dispersion, low proportion of yarn features, and susceptibility to background noise interference in visual ring spinning breakage detection in dynamic environments. Experiments have shown that the algorithm proposed in this article can meet the real-time and accuracy requirements for factory inspection.

The algorithm currently has difficulty to identify linear noise and yarn features, as the linear filter only considers its morphological features and does not combine its color and morphology. Neural networks can solve problem effectively. In the future, the standard of industrial real-time performance can be achieved by streamlining network models and improving hardware configuration. It is necessary to study how to combine the characteristics of break detection with deep learning to achieve high-precision, robust, and efficient yard breakage detection methods.

Key words: ring spinning, yarn breakage detection, machine vision, morphological operation, threshold segmentation, spun yarn