Automated Machine Learning for Industrial Applications - Challenges and Opportunities

Industrial applications of machine learning (ML) introduce specific challenges for the automation of ML workflows. The long lifetime of equipment and plants require mechanisms to ensure the functional safety of deployed predictive models. The application of sensor platforms result in large amounts of data with high throughput, which necessitate a high degree of automation. However, functional safety is often achieved through manual model validation, model selection, and subsequent model performance monitoring tasks by domain experts. Existing tools and best practices in the area of MLOps provide support for, e.g., automated processing of data, automated execution of ML tasks, or feature rich user-interfaces, to aid model validation and selection. However, they still depend on human interaction in these critical phases. Recent advances in ML allow for the integration of prior knowledge in the ML training phase. We identify this development as a crucial prerequisite for automated ML workflows to reduce the strain on domain experts. The integration of prior knowledge increases trust in ML predictive models, as they are guaranteed to adhere to certain pre-defined behavior. Therefore, predictive models, created by such ML algorithms, do not require domain experts for e.g., manual model validation. In this work, we identify challenges of automated ML in industrial applications, encountered in cooperations with our company partners. We highlight opportunities for industry that arise through the application of automated ML pipelines with integrated prior knowledge, that facilitate a high level of trust in the model's predictive performance without the need for manual validation.