Smart Quality Control Discover injection molding errors more cost-effectively using AI and robots

Related Vendors

FAULHABER Drive Systems

Wuxi InfiMotion Technology Co., Ltd.

Image analyses to detect errors in injection-molded plastic parts have been expensive and time-consuming up to now. Cologne researchers want to change this using Machine Learning (ML) ...

In injection molding, plastic products with visually demanding surfaces should be created—among other things, when it comes to the automotive industry and its requirements. To detect faulty parts, image analysis methods can be used, as researchers from TH Cologne point out. However, these methods must first be time-consuming and therefore expensive to train. The reason is that the error images in plastic parts are very diverse, as it continues. Therefore, the usability of artificial intelligence in the production of injection-molded parts is limited. In order to improve and speed up quality control, TH Cologne, together with two industrial partners, has now developed and tested an automated, camera-based AI option.

Does the injection molded part fit or not?

In plastic processing, the rapid detection of reject parts is of high interest. Especially for components that are produced in large quantities, for which injection molding is actually ideal. Particularly in the context of quality control, it is very complex to collect and verify the necessary training data during ongoing operations. The use of AI is nevertheless useful, despite all the complexity of the error images, because manual quality control takes a long time and does not always achieve the goal. The aim of the project "QuKu-ML: Camera-based quality assessment in plastic injection molding with the help of ML strategies" was to simplify the quality control of a component for the automotive industry manufactured by an injection molding machine using an algorithm.

Detect injection molding errors smartly and quickly with a 91 percent accuracy rate

The team at TH Cologne programmed a robot to automate this process by placing the component in different positions in front of a camera. This way, they say, images can be taken from 16 different perspectives. With a final dataset of about 1,600 images, an artificial intelligence was then trained to quickly recognize deviations from a defect-free component such as scratches, cracks, missing structures or deformations. For the most effective anomaly detection, the researchers then analyzed the distribution of values ​​of the 1,600 so-called heatmaps. If the number of pixels with an anomaly value is above or below a threshold, the AI ​​recognizes a deviation. This is then followed by the command to sort out the corresponding component. With this method, the experts have achieved a hit accuracy of 91 percent, it is emphasized.

When the AI knows what perfect plastic parts should look like ...

The AI-supported detection of anomalies in plastic components offers a number of advantages in an industrial context compared to usual error detection. Because the latter required a sufficient number of images on which the respective error types should be clearly recognizable. The defects must be manually marked and labeled. But perfect parts are usually available in much larger quantities compared to defective parts. So, artificial intelligence can be trained to use perfection as a distinguishing feature instead of the error. The result is an anomaly heatmap, on which defective image areas take on high anomaly values that can be represented in color.