Detection and Diagnosis of Repetitive Surface Defects for Hot Rolling Processes

This paper presents a new technique for automatically detecting and diagnosing the repetitive surface defects in hot rolling. In the paper, surface image data of the hot rolled bars is online collected using advanced image sensors, and an automatic defect detection algorithm is developed for identifying the rollinduced repetitive surface defects by integrating the Canny detector and the Hough transform. After detecting the repetitive surface defects, an estimation algorithm is further used for robustly estimating the repetitive periodicity. Furthermore, a knowledge-based root-cause diagnosis approach is proposed for identifying the stand with the broken roll failure based on the estimated period of defect patterns. A realworld case study is provided in the paper to demonstrate and validate the proposed method. INTRODUCTION Imaging-based automatic online inspection of surface defects has raised increasing research interests in recent years for the process control in hot rolling. Some fast algorithms have been previously developed for online defect detection and identification of individual defects (Jia, 2004; Li et al., 2007). However, previous research has mainly focused on detecting “individual” defects with the specific shape pattern, such as seams. Few were found on identifying “group” defects with repetitive patterns. It is known when a roll failure happens, a group of the repetitive defects will be generated on product surfaces, and thus leads to severe product scrap. In another aspect, the periodicity of the repetitive defect patterns is closely related to the specific stand where the roll failure occurs, and provides the physical mechanism to identify the root causes of the repetitive defects for a quick process correction decision. Therefore, it is highly demanded in hot rolling manufacturing industry to develop an effective monitoring and diagnostic system that can not only quickly detect surface defects, but also identify the repetitive nature of the detections for root cause diagnosis. The objective of this paper is to develop an effective monitoring and diagnostic method for quickly detecting and diagnosing repetitive surface defects caused by roll failure. The remainder of this paper is organized as follows. After this introduction section, a new detection and estimation algorithm will be introduced for identifying repetitive surface defects and estimating the period of the repetitive defect patterns. Afterward, an engineering knowledge based root-cause diagnosis approach is proposed for identifying the specific stand with the roll failure. A real-world case study is then provided to demonstrate and validate the proposed method. Finally, a summary and some future work are given. DETECTION OF REPETITIVE DEFECT PATTERN The first step of this research is to automatically detect the repetitive defect pattern. FIGURE 1 describes the major procedures proposed in the paper. First of all, the image data will go through some necessary image preprocessing steps like image cutting and vibration pattern removal. The next important step is to find the members of the repetitive pattern based on the Canny detector and the Hough transform by taking advantage of the common properties in the shape of the repetitive pattern members. Finally the detected defects will be grouped and be indentified if they satisfy the rules for checking the repetitive patterns. In the following subsections, the major steps for developing the detection and grouping rules will be discussed. Detection of Individual Defect Members For detecting the repetitive defect patterns, the first step is to find all the candidate members of the suspected defects in the images collected by the on-line imaging system. After that, we can use grouping rules to check whether these defect members have a potential repetitive pattern. To understand the unique properties of the repetitive defect pattern, it is necessary to understand the mechanism of online imaging system from where the product surface images are obtained. The core sensor is capable of capturing the objects’ surface condition at a high temperature environment. An example of image with repetitive defects pattern is shown in FIGURE 2. FIGURE 2. A REPETITIVE PATTERN EXAMPLE. In practice, 80-90% of defects induced by roll failures exhibit marks as thin horizontal lines. Therefore, based on the shape characteristic of the repetitive defects, the objective of detecting individual suspect is transformed to detecting the vertical line-shape defects. Starting from the idea of taking advantage of the shape properties, the line detection method is developed by integrating the Hough transform (Ballard, 1981) and Canny edge detector (Canny, 1986). Comparing to the detection of defects of a general shape, the proposed detection algorithm requires much less computational effort (Illingworth, 1988) and thus is feasible to be implemented for the online purpose. Moreover, the falsely detected defects due to complex background noises are also significantly reduced. Canny Detector. The first step is to detect the edge points by using appropriated edge detector (Gonzalez, 2007). The Canny detector is selected because it is an optimal edge detector, Data pre-processing Image data Canny detector Grouping rules Hough transform FIGURE 1. DETECTION METHOD PROCEDURES. Vertical (rolling direction, y-axis) Horizontal (x-axis) which has a low error rate and the good edge point localization capability to generate thin edges with respect to other derivative methods (Sobel, 1970; Marr and Hildreth, 1980). The implementation of canny detector is given as follows: The image object can be denoted as ( , ) f x y as shown in FIGURE 3(a), where the repetitive defects are marked. and x y are the coordinates of image pixels. The image ( , ) f x y will be firstly smoothed by convolving a 2-dimensional Gaussian function in order to avoid the effect from noisy pixels. After de-noising, the image is denoted as ( , ) I x y , which are gradient magnitudes having ridge thinning. To find the real edge points in ( , ) I x y , the double thresholding strategy is used as follows based on two defined thresholds: i.e., the low threshold L T and the high threshold . H T If the points satisfy  ( , ) H I x y T , (1) these points are called “strong” edge points and identified as edge points directly. If the points satisfy   ( , ) L H T I x y T , (2) these points are called “weak” edge points. If the weak points are the neighbors of the strong points, they are identified as the edge points; otherwise they will be ignored. If the gradient value range is scaled to 0 and 1, the high threshold H T can be selected from 0.01 ~ 0.3 based on the image conditions from different plants. The low threshold L T is suggested to be 30% ~ 50% of . H T Finally all the strong points and the corresponding neighboring weak points will be detected as edge pixels. The output of the Canny detector, as shown in FIGURE 3(b), is a binary image ( , ) b f x y , i.e., 1 ( , ) is the edge point; ( , ) 0 otherwise b x y f x y     (3) Hough Transform. After using the Canny edge detector, the original image is transformed to a binary image. On the binary image, the Hough transform (Ballard, 1981) is used to identify those defects having a line shape and consider them as suspect members of the repetitive defect patterns. In the Hough transform, the following representation of a line is used: FIGURE 3. (a) THE ORIGINAL IMAGE WITH REPETITIVE PATTERN; (b) THE BINARY IMAGE OBTAINED AFTER USING CANNY DETECTOR.      cos sin x y (4) The implementation of the Hough transform is given as follows. Firstly, divide the values of  and  with equal steps within the parameter ranges of   min max [ , ] and   min max [ , ] . For  , the range of       5 5 is used to identify horizontal lines. Let I and J be the numbers of subdivisions for  and  , respectively. Given an edge point ( , ) x y found in the Canny detector, for each i ( 1, , i I   ), calculate the corresponding value of  using equation (4). The calculated  value is approximated to the closest subdivision value  , j (   1 j J ). Denote ( , ) A i j as the accumulator value corresponding to the accumulator cell (  j ,i ). When a pair of (  j , i ) is found using the searching method discussed above,   ( , ) ( , ) 1. A i j A i j This step is iterated until all edge points have been checked. Afterwards, find all the accumulator cells that satisfy:  ( , ) a A i j T (5) where a T is the threshold to measure points’ concentration on a line. The lines corresponding to the found accumulator cells are detected, as shown in FIGURE 4(a). FIGURE 4. (a) THE DETECTION RESULT AFTER HOUGH TRANSFORM; (b) THE FINAL DETECTION RESULT AFTER USING TWO RULES. Two Rules for Identifying Repetitive Patterns Two rules, which are called grouping and identification rules, will be presented in this section, in which the grouping rule as Rule 1 is used to remove the falsely detected defect members based on the locations of defect lines; and the identification rule as Rule 2 is used to further reduce false alarms based on the lengths of defect lines and to check whether a repetitive pattern is existing. The flowchart of these two rules is illustrated in FIGURE 5. FIGURE 5. PROCEDURES FOR IDENTIFYING REPETITIVE DEFECT PATTERNS. Rule 1 is used to group the horizontal lines based on their locations along the horizontal direction because repetitive defect members should correspond to the same failed roll, which generates the repetitive defects having the overlapping location along the horizontal direction. For this purpose, each line is represented as a location vector , l ( ) T L R l l  , where L l and lR are the x-axis (horizontal) coordinate corresponding to the left point and right point of the line, respectively. Before using Rule 1, all horizontal lines are sorted by L l in the ascending order as   (1) (2) ( ) l ,l , ,l , l m M . Beginning from the most left one (the one with the smallest x-coordinate (1) l ), if the location vector of he selected line ( ) l m satisfies ( 1) ( ) m m R L l l l T    , (6) ( ) l m will be grouped into the previous group with  ( 1) l m ; otherwise ( ) l m will be put into a new group. Rule 1 will be iterated until all the lines are checked. l T is the threshold, which is used to measure the overlapping length of two lines along the horizontal direction. Rule 2 is used to remove the abnormal lines that have either too large or too small lengths by comparing the lines within the same group. For each grouped lines, the mean u and the standard deviation u std of the line lengths are calculated. The line whose lengths are out of the interval [ 3 , 3 ] u u u std u std   will be removed. After using Rule 2, if there are more than three defect lines are identified in the group, that group will be considered as the repetitive pattern group, which are shown as FIGURE 4(b). ESTIMATION OF DEFTECT PERIODICITY FOR ROOT CAUSE DIAGNOSIS