In the first part we saw how the Intelligent Operations Platform can speed up the defect detection process, using the real-life example of an automotive manufacturer who produces injection molding parts that are later used at the vehicle assembly line, dramatically shrinking error-resolution time. Now we will look at how we can eliminate the root cause.
Moving to the source
In order to detect the root cause, cameras are installed upstream at each of the five machines that produce this part. As computer vision is applied, the already-trained model can be reused right from the start (we call this transfer learning). The settings from the former location are now transferred to the five new locations without manual intervention. The system is continuously improved with the images captured during process execution. The reduced cost of equipment compared to conventional machine vision makes it possible for all the machines to be equipped with the cameras simultaneously. There is no need for people to monitor the situation. The massive data input from all the machines improves the model within hours. It soon turns out that the defective parts are introduced at one machine only. This raises new questions, as all the machines operate in the same hall under the same conditions.
Taking a detailed look
To get as many details as possible, the production engineer sets the quality check to a “stricter” mode. A fine-tuned model now detects even minor, non-reject deviations. This is not required in normal run mode and would create too many pseudo-errors.
Assisting hypothesis checking
To check for external influences, the experts want to record structure-borne sound. They attach a sensor at the machine and connect it to the IOP. This allows the experts to map both machine and external data to the defect detections. It turns out that a characteristic vibration of the machine corpus occurs every time the defective parts are produced, however the source of this vibration is still unknown. The production engineer defines process parameters he assumes are relevant to the problem.
Knowing when things go wrong
To get more insights, the production engineer then sets up an alert. The alert notifies him on his mobile every time the problem is detected. In addition, the parameters are explicitly marked in the continuous parameter recording stream. This allows the engineer to easily compare the parameter values for different situations.
Fixing the error
The compiled parameter comparison provides evidence: The defects are introduced while the mold is being heated. An irregular heat profile causes friction when the mold is closed. This friction is detected by the structure-borne sound detector. In this situation, the defective parts are produced in one of the three cavities of the mold. The engineer fixes this issue by modifying the heating parameters and the defects are gone.
With this customer scenario we have described how IOP can shrink error-resolution times dramatically. Using the acquired data our customer can continuously train improved models which is giving him a competitive edge.
Find out how Capgemini can accelerate the effectiveness of your digital manufacturing operations at scale with our “Factory of the Future” approach here:
Read Part 1 of this blog series here.