THANK YOU FOR SUBSCRIBING
In the production of premium vehicles, individualisation and quality play a key role. The current BMW 3 Series, for example, is available in more than one billion possible configurations. At the same time, the integration approach of manufacturing vehicles with a wide variety of drive systems–petrol, diesel, plug-in hybrids, and battery electric vehicles–on the same assembly line increases the complexity of the production system. In addition to manual visual checks or conventional, pixel-based computer vision methods, Artificial Intelligence (AI) offers an innovative, high-performance method for quality assurance in production.
The technology of Deep Learning, a method of Artificial Intelligence, is particularly suitable in image processing for the tasks of classification, object detection, or semantic segmentation. In the application of AI, a distinction is made between the training phase and the subsequent phase of live operation, the “inference”. In the case of image processing, an initial data set of photos is required for training. The pictures must be arranged in such a way that relevant environmental parameters such as positioning, orientation, focusing, illumination, reflections, etc. vary in a targeted manner so that they approximate the scenarios to be expected during the inference. In addition to the photos, so- called labels are required that reflect information about the type and location of relevant objects that is adequate for the task. Photographs including labels are then used to modify the internal weights of a deep neural network in such a way that it matches the labels in the training data as accurately as possible.
In order to efficiently achieve a powerful model, the so-called Transfer Learning is applied. In this process, an underlying neural network is modified, but not completely learned from scratch. Using an existing network architecture with given weights, the modification is carried out with the help of a framework and a new training data set. Typically, this procedure requires a training data set of a few hundred photos.
The support of the employees is fundamental to all innovations in the BMW Production System. This applies equally to AI systems. Compared to an inspection of certain quality features by employees or so called camera portals, which are based on conventional methods such as pixel-based grey scale comparison, AI offers significantly better performance on the one hand, i.e. fewer pseudo errors. On the other hand, AI works robustly under essentially all production conditions. The BMW Group, for example, has implemented the inline detection of model lettering in assembly or the examination of sheet metal parts in real time using AI. To ensure that all production employees experience the benefits that AI can bring to their daily work, these and other successful AI applications were presented in a broad internal communication campaign.
"AI solutions in the BMW Production System are designed so that employees can use them immediately–with little to no configuration needed"
In the meantime, numerous AI applications in assembly, press shop and logistics are running in the production sites of the BMW Group at locations worldwide. Further AI applications are currently being developed in all technologies, each of which is developed and stabilised exclusively at one location before the worldwide roll-out takes place. The added value for employees manifests itself primarily in the following aspects: Foremen are unburdened from pseudo-errors. Because AI achieves a significantly better assessment and in many cases 100 percent robustness, foremen no longer must double-check issues or examine pseudo faults in quality systems. The direct production employees, benefit from more appealing work content, and another method of quality inspection: Instead of repetitive, fatiguing control activities, they can concentrate on value-adding tasks and process improvement. Overall, the awareness of quality increases if all categorisations of parts or components are correct.
The central approach is to bring AI to the people and offer self-service for production employees. This means that the AI solutions in the BMW Production System are designed so that employees can use them immediately–with little to no configuration needed. In order to democratize Artificial Intelligence and make it available to a wider range of users, the BMW Group even goes one step further and publishes these algorithms as open source. The published algorithms include programming interfaces, so-called APIs (application programming interfaces) and the labelling tool developed in the Innovation Lab of the BMW Group IT. With the help of the APIs, it is possible to efficiently use established frameworks: APIs for both training and inference have been published. These allow a particularly simple connection of the frameworks.
AI allows the efficient and at the same time robust design of quality inspection systems. It provides a much more robust recognition than established manual or automatic methods. At the same time, it enables operation under almost any environmental conditions. It thus supports the complexity management in production. In the short term, it is useful for significantly shortening quality control loops.
In the BMW Group, technology serves people. In production, this means that Artificial Intelligence can be used intuitively by all employees–without them having to be data scientists, without having to program a single line of code. The following aspects contribute significantly to the democratisation of AI: AI solutions must be designed in such a way that they can also be set up and used by users without in-depth IT expertise. In production, this means that AI extends the toolbox. Technology must not be an end in itself. People are always the final decision-makers–they can decide for themselves which innovative solution can best support them in a given task.