Bosch Pilots the Use of Synthetic Data to Reduce Time Needed to Develop AI Applications

Bosch, the well-renowned German engineering and technology company, announced just before the end of 2023 that the company was piloting generative Artificial Intelligence (AI)-based projects at two of its plants in Germany (Hildesheim and Stuttgart-Feuerbach). Primarily the pilots’ objectives are to investigate whether generative AI can help the manufacturing teams increase productivity and efficiency with the use of synthetic data (data created by an AI model attempts to replicate the nuances and patterns of the real-world data on which the model is being trained. It is used to save time generating models when real-world data are scarce).

The year 2023 saw many Industrial and Manufacturing (I&M) companies looking to understand how generative AI can improve their operations and experimenting with the technology see (ABI Research’s Generative AI Use Cases in Manufacturing presentation (PT-2763)). Bosch showcases the potential for using synthetic data, but I&M companies need to be aware of the potential pitfalls of such an approach.

Bosch is already making extensive use of AI to support and improve its manufacturing operations. Currently, 50 plants use AI to identify potential breakdowns on the production line. Furthermore, 20 plants use AI to underpin their machine vision inspection processes to identify scratches and other imperfections. Staff at the Hildesheim plant use AI for production scheduling, while the Stuttgart-Feuerbach plant incorporates AI in its processes for testing components.

The step change that generative AI provides is the use of synthetic data to support the processes by augmenting the real-world data. At the Hildesheim plant, synthetically generated images have been used to support staff training by generating standardized images of the complex production systems for electric motors. Previously, staff at the Stuttgart-Feuerbach plant were unable to develop either rule-based or AI-assisted optical inspection processes to support the inspection of the fuel-injection components due to the complexity of the products and differences in the production processes. The generative AI model can accommodate the different processes and product arrangements to recognize errors thanks to datasets received from equivalent production lines across the company that contribute to developing synthetic images. The AI solution will inspect the components and notify inspectors on occasions when it’s unsure.

Bosch predicts that the time required to plan, collect data, structure the models, and then introduce AI-based applications will decrease from 6 to 12 months to a matter of weeks. If successful, all 230 Bosch plants will be able to develop synthetic data to support their individual use cases.

The use of synthetic data is the next chapter in Bosch’s Industry 4.0 strategy. The company has been digitizing assets and using data analytics for over 10 years; approaching half the company’s plants incorporating AI to support their operations. The hope is that the use of generative AI will enable the plants to operate more efficiently both from productivity and environmental perspectives.

The collection of a sufficient volume of images and data can delay the progress of AI-based use cases. The use of synthetic data seeks to plug that gap. However, project teams considering developing AI applications based on synthetic data should be mindful of the following issues:

• Data Quality Still Applies: Project teams must ensure that the data/images generated still preserve the unique characteristics of the real-world context on which the model will be trained. The model shouldn’t be prone to hallucinating on the synthetic data.
• Ignore the Tantalizing Time Frames: Quality control, as in many other aspects of I&M operations, needs to be at the forefront of implementers’ minds. In addition, once you are happy with the model, you cannot just set it and forget it. Implementers need to check that the model doesn’t become a closed system with little input from the real-world data.
• Apply a Cynical Mindset: Synthetic data promises much, but users need to regularly scrutinize, augment, and annotate the data to check for biases and to validate the model’s accuracy regularly.
• Start with a Low-Risk Use Case: Implementers unfamiliar with the use of synthetic data should identify low-risk use cases to start with and develop their processes for ensuring the model is performing as designed. Doing so will not only build expertise, but will identify underlying data collection processes that the company needs to improve before tackling more complicated and critical use cases. Furthermore, by building expertise gradually, project teams will have a better understanding of the benefits in terms of time savings, productivity gains, etc. that a deployment can deliver.

Sometimes, implementers would be better suited to examining their processes for collecting high-quality data as opposed to looking at synthetic data from the outset.

Bosch is certainly an I&M company to track closely, both as an example for other implementers to follow and because of its adoption of the latest technologies. Bosch is one for suppliers to target as well. The company has been acknowledged as a leading firm with regard to Industry 4.0, with four plants recognized as part of the World Economic Forum’s lighthouse factory, including two its work with AI.