The displays of future vehicles are likely to be more advanced and sophisticated than what we see in cars today. In the same way, production conditions of the same displays are becoming more complex and demanding, specially the quality control and the final inspection that represent the last opportunity to see the quality and condition of the shipment, while it’s still in the factory. This is being performed manually with all the drawbacks it entails. The main goal of this work is the development of a fully automatic final control inspection system dedicated to the displays clusters. The developed system consists of an automatic machine integrating an artificial intelligence model to increase detectability of defects. To construct the model, the machine permits a first double visual inspection: one manually performed, and another performed by a visual inspection system. This one is composed by high resolution cameras to capture images of the display, and a pre-trained AI algorithm that analyze the images. Then it is possible to use the new images to retrain the model. That is, the machine can continuously improve its ability to detect defects, as it is exposed to more and more data. This will allow the machine to become increasingly accurate and reliable over time, until reaches the target specifications and will become totally automatic. To facilitate and accelerate this model training, the double inspection is performed in parallel using a rotary indexing table with two jigs and a dedicated labeling software. The chosen automatic visual inspection system permits the capture of all the defects that are present in the collected samples, even the smaller ones with 0.3mm in diameter. The labeling software permits the visualization of captured images to annotate the defects and classify them into one of the three classes: dirt, scratch and pin hole. The preliminary results of the trained model with the collected samples are also satisfactory to be applied in the line production. This dataset is composed by 34 displays, 2 of them with no defects. But as the machine learning model treats the images by patches, in fact there are many more samples available to train and validate the model. The development of Artificial Intelligence Inspection System to be used in an industrial line production has been difficult, mainly due to the lack of data available for training AI algorithms and the need for the images to be analyzed in real-time, requiring powerful computing resources, and thus increasing a lot the cost of the solution. Furthermore, when going to the industrial environment, the machine should be ready to produce quickly, so there is not much time to perform the double inspection and optimize the model. The system under development takes into consideration the limitations described above and includes some tools to mitigate and accelerate the process of train an artificial intelligence model, such as the double visual inspection, a tool to annotate the captured images and a method to retrain the model locally. In that way, the machine will be ready to produce when installed in a real production environment because of the manual inspection, and in a short time become totally automatic due to its ability to learn and adapt. The development of an Automatic Final Control Inspection System for Automotive Displays has the potential to significantly improve the quality control of automotive displays. By using an artificial intelligence model and high-resolution hardware vision, the system can detect and identify any defect in the displays. However, the implementation of such a system poses several challenges, including the lack of real images to train the models in advance, real-time processing, and the cost of the solution. Addressing these challenges is crucial for the successful implementation of an AI-based inspection system for the final quality control of automotive displays.

Dipl.-Ing. João Queirós, RDI Manager, Controlar