Research On Tundish Slag Feature Extraction In Steel Level Measurement

In steel metallurgy industry fields, steel level measurement in continuous casting tundish has been trapped in a problem:atrocious measuring circumstance caused by high-temperature of molten steel and the flux slag on upper layer make it difficult to measure molten steel level accurately. According to this fact, a new method solved by monocular vision is put forward to molten steel level measurement based on temperature field distribution. However, in the actual measurement process, due to slag crust easily, appear the situation of completely obscured vision or damage to the temperature characteristics, leading to measurement error. So for this case, do some research on slagging image de-noising, edge detection, feature extraction and classification.First of all, research on slagging image pre-processing. As the site of the metallurgical complex environment, making the measured characteristics can not be reliably extracted. In order to improve the stability and accuracy of measuring system, this paper uses the multi-resolution character of wavelet transform and proposes an improved wavelet threshold de-noising method. As the adaptive processing based on the relation of the signal variance coefficients and the signal variance coefficients, it can hold the real signal while de-noising. In addition, the measured target is often abnormal in the context of interference, superposition of a large number of false information, which affects the reliability of feature extraction. In order to reduce interference, this paper extracts the ROI first, and then in the ROI, use the gray increased to make the characteristics of the signal output maximization.Secondly, study on slagging image feature extraction. According to the gray distribution characteristics of typical image, research on the accurate edge extraction algorithm of the measurement bar, and propose a edge detection quick algorithm which suits industrial site. According to the geometric contours extracted, define six characteristic to describe the target information. According to the statistical characteristics of gray, define four characteristic to describing the image texture information.Thirdly, Research classification of slagging based on support vector machine. Characteristics of the raw data collected normalized pre-treatment, and simple to use PCA for feature10-dimensional feature dimension reduction for the5-dimensional, to speed up the model of training speed and improve the classification accuracy. RBF kernel used in this experiment, grid search strategy through the best search parameters, classification results:the accuracy rate of training samples:87.65%; The accuracy of the test sample:80.12%; The average classification accuracy of85%which meet the needs of field measurements.Finally, using Tundish slag detection technology in the industrial field, analyze and compare the before and after application of the results show:in level measurement process, a timely and accurate detection of slag increases the stability and robustness of the measurement system.