67 based on the latest scientific standards. The application case was the component manufacturing of a gearbox. The employee is presented with the two halves of the gearbox on a turntable (Figure 1 and 2). Their task is to insert various components before the turntable transfers the correctly equipped gearbox halves to a robot. In one half, the employee must insert an adjustment disk and a flat gasket, while in the other half, they must install 18 screws into the designated holes. Two additional holes, which are in the same plane and visually barely distinguishable from the other 18, must remain unfilled, as the robot arm will grip the gearbox half in the next process step. Possible errors include failing to insert the adjustment disk, flat gasket, or screws, or mistakenly inserting screws into one of the unassigned holes. Such errors can lead to high follow-up costs and machine downtimes. The goal of the inspection system from Fraunhofer IOSB was to identify such errors. It is deployed after the employee has inserted the components and given the command for the next process step (transfer to the robot) via the control system. The automatic procedure checks based on vertical images of the gearbox halves whether the components have been placed correctly (inspection result „i.O.“ for „in order“). If not (inspection result „N.i.O.“ for „not in order“), the inspection results are forwarded via the SPS to the plant system, and anomalies are displayed to the employee so they can correct the errors. Camara arrangement and data processing of the Fraunhofer system comply with the strict regulations of Porsche regarding the protection of personal data. The underlying procedure for anomaly detection is based on a machine learning model known as the Patchcore model. This model performs one-class classification („1-class classifier“), learning only the class „No anomaly present.“ During the training phase, only positive examples of the inspection situation are presented to the algorithm—specifically, images of the two gearbox parts with correct assembly. Thus, the model learns only what the desired state looks like. A significant advantage is that positive examples are often readily available from normal operations, as employees typically work correctly. Anomaly examples occur rarely and would therefore be significantly more expensive to acquire. When deployed, the trained Patchcore model produces an anomaly score for each presented image and a heatmap indicating all areas where an anomaly has been detected. As with all classification decisions, it is possible for this to yield incorrect results, especially when the situation is near the decision boundary. To communicate a decision boundary during employee training in quality control, companies use so-called boundary patterns: Employees are given objects or images that are evaluated as intact and those that are already deemed not intact. In borderline cases, the employee is required to choose one of the evaluation classes. The automatic decision of the inspection procedure also uses a decision boundary, which is individually adapted depending on the application case. As with humans, the automatic procedure can also yield false negatives, failing to detect an anomaly that exists („Pseudo i.O.“), or false positives, detecting an anomaly that does not exist („Pseudo N.i.O.“). Both are to be avoided, as a „Pseudo i.O.“ leads to defective gearboxes with follow-up costs in the process and machine downtimes, while a „Pseudo N.i.O.“ results in time delays for the employee, who unnecessarily needs to verify a correct assembly. To demonstrate the applicability and benefits of the procedure, Fraunhofer IOSB has implemented a setup on the production line at Porsche. A camera hangs over each gearbox half. When the employee presses the button for transfer to the robot after completing the assembly, each camera takes a picture, which is then analyzed by the inspection system. If the classification result is „i.O.“ (no anomaly), the transfer occurs. If the classification result is „N.i.O.“ (anomaly), the employee receives a prompt to correct the issue. To showcase the potential of the procedure, it ran in the background for two weeks during work shifts. For the employees on the production line, there was no change to the status quo, as the indication function of our system for the classification result „N.i.O.“ was not activated. Porsche AG stipulated that the Pseudo-N.i.O. rate must remain below one percent for certified usability. A higher rate would significantly delay employee tasks and undermine trust and, consequently, acceptance of the automatic procedure. Analysis of approximately 1,000 recorded images of each gearbox half yielded the following results. There was no occurrence of „Pseudo i.O.“ (an undetected anomaly). Two „N.i.O.“ cases were detected, one in which a screw was not installed and another where a screw was placed in an unauthorized hole. The Pseudo-N.i.O. rate for the screw installation and the insertion of the adjustment disk was 0.2 percent. Thus, the requirement of a maximum one percent Pseudo-N.i.O. rate was significantly exceeded, demonstrating the usability of the inspection system. Fraunhofer IOSB www.iosb.fraunhofer.de/en.html Department Human-AI Interaction https://t1p.de/1buv6 Overall view of the production station with gearbox halves and cameras on the ceiling structure.
RkJQdWJsaXNoZXIy MjUzMzQ=