Unplanned downtime in Industry 4.0 manufacturing poses significant challenges, yet offers substantial benefits such as increased productivity, minimized defects, reduced unplanned downtime, and optimized resource utilization. It can be reduced by forecasting machine failures from Industrial Internet of Things (IIoT) sensor streams. The Bosch Production Line dataset captures this challenging setting that contains 4,264 heterogeneous station features for 1,183,165 parts and only 0.58% labeled failures, with many sparse, high-cardinality nominal codes. In this study, we develop an end-to-end pipeline that normalizes raw station measurements into a relational schema, derives 178 station, line, and path aggregates from 969 numeric columns, and compresses 2,139 nominal columns into 16 supervised Weight of Evidence (WoE) risk descriptors to enable learning under extreme imbalance. The workflow is implemented in Apache Spark on Databricks to support scalable feature engineering and low-latency scoring on streaming data. On a held-out test set, XGBoost achieves 0.966 AUC-ROC and 0.793 MCC with 0.929 F1, while end-to-end feature engineering completes in about 45 minutes and model retraining in about 30 minutes on a 6-node CPU cluster. These results indicate that WoE-based nominal compression combined with a real-time capable infrastructure based on distributed Spark processing enables production-feasible failure forecasting. Finally, we release the Databricks notebooks and Spark code to support reproducibility of the results presented in this study.
Zdravevski et al. (Wed,) studied this question.