Modular Framework for Scalable Analysis and Modeling of Time-Dependent Industrial Sensor Data Using the Example of The Bosch Production Line Performance Dataset

The processing of large multivariate sensor data from industrial manufacturing systems requires scalable and robust Artificial Intelligence methods, which focus on the temporal dependencies within the data. This paper presents a modular framework for data preliminary processing, feature engineering and modeling of time series, specially adapted to the Bosch Production Line Performance Dataset. In addition to classic time series baselines, machine learning-based models with manually extracted time series features as well as deep learning architectures such as long short-term memory and Transformer are evaluated. The implementation takes place in a distributed Apache Spark environment to efficiently address data volume and high dimensionality. Our results show that Deep Learning models better capture long-term dependencies, while classic Machine Learning baselines achieve predictions quickly, scalable due to effective feature engineering. The presented pipeline is a practical and scalable approach in terms of analyzing industrial sensor data which contributes to a data-driven transformation in smart manufacturing and Industry 4.0.