Industrial Sector

Quick and easy identification of 13,000 items via cell phone

Training a neural network for Bosch Cognitive Services to reduce errors and increase first hit rate.

  • Bosch Cognitive Services needed to improve its AI based service which identifies spare-parts via cell phone photo. The goal: reduce effort and improve the first hit rate.

  • The Zühlke team delivered a successful trained and tested new AI solution within eight weeks.

  • The new solution is far exceeding the initial goals: It simplifies the data acquisition process, saves time and resources and increases the first hit rate of 84%.

Julian Weiss Bosch Cognitive Services
' With Zühlke as our second source partner, we were able to improve our existing solution in just eight weeks. We were also able to far exceed our initial goals both in terms of time and overall accuracy of the solution. '
Julian Weiss
Technology & Innovation, Bosch Cognitive Services

Initial goal: Simplify the AI training and increase the first hit rate for users

Bosch Cognitive Services (BCS) offers a service which identifies spare-parts within seconds via cell phone: A single photo is compared with a catalogue of over 13,000 products to identify the correct item. The image evaluation is based on a customer-specific machine learning (ML) model. So far, BCS needed photos from two sources to train the AI solution: Images from an automatic scanner as well as a large number of cell phone photos. BCS relies on Zühlke as a second source partner to improve the existing solution they have been working on for 3 years. The goal: Establish an AI model that displays the right product with a first hit rate of more than 70%.

A successful tested and simplified AI model within 2 months

The team solved this challenge within eight weeks: The existing processing pipeline was examined and revised. Through appropriate normalisation, the experts simplified the AI training. The result: Cell phone images are no longer needed for validation. Compared to the existing solution, the new model is based on a multi-level approach. Training and model optimisation are implemented in the Azure cloud. After successful testing, the trained model was deployed to BCS for internal validation.

Bosch Cognitive Services staff using the mobile app photo feature

The new AI solution reduces the effort and exceeds the initial goals

With the created training pipeline, the AI training can be performed purely on the images from one source: the automatic scanner from BCS. This significantly simplifies the data acquisition process, saves time and resources and increases user acceptance. The indication of the error rate per process step in connection with an estimation of the potential of further improvement possibilities supports BCS with product improvement and investment planning together with Zühlke. And with the improved AI solution, BCS achieves a first hit rate of 84%, far exceeding the initial goals.

Bosch Cognitive Services staff inspecting spare parts