How Artificial Intelligence Makes Quality Traceable Down to Component Level

In proof images created with the shape-from-shading method, even minute faults like scratches or scoring can be evaluated. In textural images, differences in coloring or lighting can prevent a stable evaluation. (Image: Kistler Group)

Manufacturers in the punching and stamping industry have to meet demanding requirements and expand their quality control significantly. Here, technologies based on artificial intelligence (AI) have been used to achieve these goals for several years. They enable a comprehensive quality control and simultaneously improve its efficiency. The combination of AI-based software, powerful optical inspection systems and marking systems equip punched parts manufacturers for the demands of the market.

The widely practiced random inspection of components in the production of stamped parts is not only time-consuming, but also far from delivering end-to-end traceability. Yet there are solutions that can replace inspection based on random part sampling with an automated, comprehensive quality control. They offer traceable quality down to the component level. For the highest possible benefit, it is essential that the corresponding technologies – special illumination, imaging and marking techniques – can be integrated in existing machines and equipment and function well with the AI-based anomaly detection.

The Shape-From-Shading Method

The optical inspection process known as shape from shading uses special illumination and imaging techniques. It separates the information about the texture of an inspected part from its topological characteristics and makes the exact inspection of individual components in the stamping and punching industry possible. With this method, even minute anomalies in individual parts can be made visible that would remain undiscovered with other methods.

During the inspection process, the inspected part is illuminated from several directions and captured by a camera. The resulting images only show the dispersion of light and shadow. Based on these (real) single images, the software calculates different topographical images that only show the 3D information about the surface of the inspected part. This makes quality inspection independent of changes in the surface of the inspected part, for example differences in color or brightness, that would show up in the texture image and would prevent a stable evaluation. With the shape from shading process and conventional image processing techniques, even scratches, cracks or indentions with heights or depths of only a few micrometers can be detected reliably.

Due to the special LED illumination, the method works reliably even with dark or shiny surfaces. A sophisticated algorithm that compensates movement allows the process to applied to moving objects as well and therefore to be used in automated inspection technologies with high part throughput rates.

Kistler’s KLM 621 laser marking system can apply markings to as many as 2,500 parts per minute like at the production site of KRAMSKI GmbH, Pforzheim. (Image: Kistler Group)

The AI-based process is used to inspect and evaluate the images taken. It is a special technique based on deep neural networks (DNN). Here, a convolutional autoencoder works in combination with differential image generation to make unusual or unexpected deviations in images of inspected parts visible. This kind of anomaly detection shows its strength where conventional image inspection reaches its limits or requires a high degree of expertise: in fault detection in complex textures. Neuronal networks are often the only option when it comes to variability among good parts and decision criteria between OK and NOK components that cannot be sufficiently described by mathematics.

To use anomaly detection the software needs to be taught first. The deep neural network is fed with images of OK parts and learns their characteristics and the skill to reconstruct them as precisely as possible. Provided that the anomaly is easily recognizable, users can train the neuronal network with colored or monochrome images as well as with depth or curvature images.

Minimizing Pseudo Scrap

The anomaly detection method makes use of the fact that the convolutional autoencoder cannot reconstruct divergent image contents and structures. Therefore, the reconstruction of an inspected part will not include the anomaly but look like the corresponding good part. It only takes establishing the differences between input and output images to receive the actual anomalies.

The anomalies can then be classified through further image procession methods or – if needed – with the help of an additional neural network. The AI-based software triggers the separation of the part if an anomaly is detected. After a few production batches, users can feed the software with more pictures of OK parts that may have different characteristics than the parts in the first round. This allows users to further refine the AI and minimize the percentage of pseudo scrap.

With the use of AI, manufacturers can vastly improve their quality assurance regarding unusual or only sporadically occurring defects. These defects are often not covered by the mathematical parameters used in conventional, rule-based inspection methods. Therefore, parts with these faults do neither get detected nor sorted out. Going forward, manufacturers will in all probability use a combination of conventional, rule-based, and AI-based methods – because one cannot replace the other.

A stamping cell with the integrated optical shape-from-shading inspection method. (Image: Kistler Group)

Laser Marking of the Individual Components and Documentation

The precise marking of the individual parts that have been ruled correct after inspection is a central part of the efficient design of quality assurance and the extensive documentation that goes with it. This way, manufacturers can implement their quality assurance down to component level and document it comprehensively. Since the punching industry marks components continuously as they pass down the line with the marking on the fly method, the laser marking system downstream of the optical monitoring has to fulfill special requirements for marking speed. For example, the LASERmark KLM 621 marking cell used in a combination for quality assurance by Kistler is based on the latest fiber lasers. The cell fulfills the demands of an economically short marking time, the lowest possible position tolerance for the marking field and optimal contrast conditions. In the marking-on-the-fly process, the laser cell reaches a performance of up to 2,500 parts per minute. The high-precision trigger sensor by Kistler ensures a reproducible position accuracy of less than 0.01 mm. It allows the efficient and complete part traceability thanks to the seamless marking or coding of every produced part.

Next to quality requirements, standards for the level of detail and completeness of the documentations are on the rise. A comprehensive and detailed recording and storage of relevant data about the produced and marked parts are mandatory for manufacturers in the punching industry. Networking of all the essential machines is the foundation of an efficient data management and the following analysis for process optimization but also the base for complete traceability at component level.

Kistler has responded to this need by integrating a OPC UA interface based on universal machine technology interface (UMATI) into all solutions in optical inspection automation. Networking the system with the surrounding machines makes it easier to monitor the process comprehensively. The system stores the relevant documents regarding the individual parts in a database. This allows users to advance their quality inspection from a random sample data base to a comprehensive inspection and traceability of the manufactured parts. Depending on the specific application case, both machine data and inspection results are presented, analyzed and statistically evaluated in real time or stored for later evaluation. In conjunction with analysis software tools like MaDaM and jBEAM users can analyze quality, production and measurement data at any time from any location, with access from the correct instance. At the same time, they can use the data for further process optimization in production. Sources of error and unnecessary paths are eliminated.

Manufacturers can meet the requirements thanks to the exact inspection of the individual components that is continuously improved using AI. The traceability of the produced parts is guaranteed by the individual marking and documentation. Not only can they uphold the quality standards, but also maximize the efficiency of their production – while guaranteeing the individual traceability of the parts they produce.

This article was written by Stephan Bellem, Head of Engineering, Kistler Group. For more information, visit here .