Finding quality-relevant surface defects
Intelligence of optical inspection systems: an interaction of hard- and software
Surface inspection systems for quality monitoring are standard practice in modern steel mills and almost every strip processing line is equipped with them. The basic design of such systems is identical. Cameras with suitable lighting technology record digital signals from the metal strip. Hard-coded algorithms process these signals into images that an automated classifier evaluates and decides, with some uncertainty, exactly what the system might have detected. In addition to reliable detection, the high accuracy of the classifier is a basic requirement for successful quality assurance. At first glance, intelligent software should perform this task to our satisfaction. However, reality often teaches us otherwise - much to the displeasure of system users - no matter how intelligent the software appears to be.
"Critical or non-critical? It depends on the type of defect."
"We have to become more efficient" has for many years been a frequently heard phrase on the management floor of the manufacturing industry - long before the energy crisis. Systems for assessing the efficiency in the production process and ultimately the product quality are enjoying sustained high demand. An example of such systems is shown in Figure 1.
Figure 1: Camera-based surface inspection system with corresponding LED lighting technology on a galvanizing line.
Source: Dr. Schenk Industriemesstechnik.
This system records the galvanized steel strip over its entire surface and specifically searches for quality-relevant surface defects. If the system detects a defect, it generates a high-resolution image which is evaluated by intelligent software. Roughly speaking, such defects can be divided into three categories which, depending on their type and size, are then assessed as non-critical or result in an entire coil being blocked for delivery or further processing. On the one hand, there are the typical contrast defects such as soiling, emulsion spots, oxide residues and many more. This type of defect is rather irrelevant in the majority of production plants, as it disappears in the subsequent processing steps, and therefore only plays a role in the last step of the process chain or in packaging. Nevertheless, these defects occur frequently in all production steps and must be distinguished from the two other, quite critical defect types.
These are, on the one hand, topographical defects such as dents, bumps, open shells and, on the other hand, scattering defects such as scratches or other damage in the metal strip. These defects do not disappear in the subsequent processes and will be visible in the final product. After all, who wants dents in their new car or scratches on the design elements of their kitchen?
Significant advantages through early detection
In the spirit of "we must become more efficient," it is important to sort out defective material as early as possible in the production process. On the one hand, it can then be used for less quality-relevant products such as construction steel, and on the other other hand, the downstream processes which are usually very energy-intensive, are not burdened with inferior material from the outset. Another significant advantage of early detection of critical defects is the ability to intervene in what appears to be a faulty process, such as a damaged roll that leaves a dent in the material with each revolution. The earlier a process defect is detected, the less material ends up in the scrap container at the end of the day. To achieve this, it is essential to correctly classify the defects found, i.e. to classify quality-relevant defects as accurately as possible and distinguish them from less critical defects. A whole range of software is available for this purpose, from the simple application of fixed rules to neural networks, which operate under the umbrella term "artificial intelligence".
It takes two intelligent components
But: Any software can only work with the information content of the image generated by the system. There is nothing more, regardless of how intelligent the software is. At this point, it is best to ask how a human being tries to evaluate a defect that he holds in his hand as an A4 sample. He will tilt and turn the sample in his hand in front of his eyes, looking for the angle from which he can best evaluate the defect. No software can do that, but fortunately there is intelligent hardware. Figure 2 shows four different arrangements, each with a different angle of incidence of the illumination. In this way, the camera is presented with four different angles of view of the material, similar to the intelligent method of the manual inspector. The result is not just one image of the same defect, but four. This helps the system user recover his habit of looking at patterns from different angles in an automated system and increases confidence in the reliability of the inspection system. Automated software can draw on the information content from four images and does not have to make a decision based on just one image.
Figure 2: Schematic representation of the four optical arrangements, here to be understood as viewing angles: Bright field directional, bright field diffuse, dark field in the machine direction, dark field transverse to the machine direction.
Quadruple information content
Figure 3 shows how an arrangement consisting of four viewing angles can be implemented in a single system. By sequencing the illuminations over time to match the scan frequency of the camera, a real-time image from four different viewing angles is guaranteed. What exactly does such a method contribute to improved discrimination of critical and non-critical defects?
Figure 3 shows how the four viewing angles can be combined in one system.
Source: Dr. Schenk Industriemesstechnik.
Figures 4 and 5 show examples of critical and non-critical defects, each of which can only be distinguished using one of the four available viewing angles. It is easy to imagine what would happen if these defects were reversed. For example, because only one bright field diffuse and one dark field in machine direction are available for material observation. Then dents and scratches end up in supposedly high quality material, while defect-free material is blocked for further processing or even scrapped. This is extremely counterproductive for efficiency, especially in the metal industry, which is already being hit by energy prices. In such a case it doesn't even matter from which viewing angle.