VIMA: Modeling and Visualization of High Dimensional Machine Sensor Data Leveraging Multiple Sources of Domain Knowledge

RfX: A Design Study for the Interactive Exploration of a Random Forest to Enhance Testing Procedures for Electrical Engines

“…In the case of engineering design, efforts have been made to visualize in‐car communication networks [SIB*11, SFMB12] or the exploration of multi‐criteria alternatives for rotor designs [CMMK20]. Recent studies have been carried out making use of anomaly detection to detect error‐prone produced parts in test stations of large‐scale manufacturing processes [SMF*20, EJS*20]. While some of the mentioned studies acknowledge the need for interpretability of applied ML models [EJS*20], considerably less work has been dedicated to the development of systems that enable the analysis of such models.…”

“…Recent studies have been carried out making use of anomaly detection to detect error‐prone produced parts in test stations of large‐scale manufacturing processes [SMF*20, EJS*20]. While some of the mentioned studies acknowledge the need for interpretability of applied ML models [EJS*20], considerably less work has been dedicated to the development of systems that enable the analysis of such models. Grounded on previous findings, we did build such a system, that visualizes the decision‐making process of an RF classifier.…”

“…Detailed information about the feature extraction process can be found in Eirich et al . [EJS*20]. The two datasets are briefly described in detail.…”

“…With the fast up‐ramping of new assembly lines for electric vehicles across the globe [HKB*19], this problem has gained much attention across industries recently. To detect faulty parts, engineers typically rely on recording and analysing signal measurements, as widely used in many other engineering domains [SMF*20, EJS*20].…”

“…RFs are relatively easy to implement, less prone to overfitting than similar approaches such as DTs, and well‐suited to detect dependencies in their feature subspace. They also serve well to find combinations of features from measurements that improve part testing, making them a good choice for our target domain [EJS*20, PBLG19]. However, without proper visual interfaces, RFs can be hard to interpret for users, specifically, for those who have little or no ML background.…”