Zeiss uses getML to improve predictive maintenance models by several percent

ZEISS is working on delivering highly reliable machines and support processes. GetML allows building new predictive maintenance applications in a fraction of the time.


ZEISS is an internationally leading technology enterprise that operates in the fields of optics and optoelectronics. In 2018, the company employed 29,000 people and had an annual revenue totaling more than 5.8 billion euros. ZEISS comprises a division called ZEISS Digital Innovation Partners which employs around 50 developers, data engineers & scientists. Here, ZEISS is at the front line of digital transformation and promotes beyond state-of-the-art solutions in the rapidly changing fields of IoT, Augmented Reality, Smart Connected Products and Customer Success.


ZEISS relies on the outstanding performance of its highly complex operating production machines to guarantee the excellent quality of its customers’ products. However, machine performances decline over time if untreated and cause economic damage. Hence, ZEISS Digital Innovation Partners developed a predictive maintenance algorithm to determine those production machines likely to fail soon. Subsequently, the system prioritizes these machines for an off-schedule maintenance inspection. The in-house-built model already worked well. Yet, identifying the respective machines is like searching for the notorious “needle in a haystack.” The question remained whether the predictive accuracy of the maintenance models at ZEISS could be optimized. Therefore, ZEISS Digital Innovation Partners collaborated with the getML team: The cooperation should show whether automatic feature extraction can help to build more accurate prediction models.


Machines quickly gather hundreds of gigabytes of raw data that ML algorithms need to be able to process. Existing ML algorithms require significant manual extra work to process and to make predictions based on data of such complexity. Current ML algorithms can’t uncover relevant features in multivariate time series that span multiple tables. Hence, a data scientist must work on feature identification and extraction for every new project manually. This procedure requires a profound knowledge of the underlying processes that cause machines to fail (manual feature extraction usually takes 2-3 months of full-time work and hence consumes significant amounts of a project’s budget). Features missing in the development of ML algorithms can jeopardize its prediction accuracy. Underperforming models then may lead to redundant maintenance services or even to the outage of production lines.


Out-of-the-box ML algorithms do not understand relational and multivariate time series that span multiple tables - getML does. Hence, getML was able to build predictive analytics models on 300 Gigabytes of preprocessed sensor data that the production machines of ZEISS stored in different files and tables. Our team closely collaborated with the data scientists working at ZEISS: In a four-week time frame, we wanted to find out whether getML could exceed the in-house-built predictive maintenance system of ZEISS in terms of speed and accuracy. Within two weeks, we developed and trained the respective model: the getML team delivered a prediction model that indicated more accurately which production machines were likely to suffer a particular defect within the next 90 days. As a consequence, getML was able to further reduce the expected downtime compared to the conventional machine learning models that rely on manually-generated features. In addition, the overall time that the data scientists spent on building similar predictive maintenance models got reduced significantly. Freed-up data science capacities can now be reinvested in the optimization of existing digital innovation products or in the exploration of new business cases. ZEISS continues to use getML in the evaluation phase of new digital innovation products.


The collaboration delivered a top-performing solution: getML’s predictive algorithm outperformed the prediction model based on manually-generated feature extraction within two weeks. The getML prediction model showed higher predictive accuracy than the already existing models. Therefore, scheduled maintenance services optimized the efficiency of service teams. Overall, getML will empower ZEISS Digital Innovation Partners to increase its project efficiency further. Zeiss can invest the resources saved by automated feature engineering in new business cases and in the development of further digital innovation products. getML continues to be the tool of choice for new digital innovation products in the evaluation phase.