Adapted Random Survival Forest for Histograms to Analyze NOx Sensor Failure in Heavy Trucks

Z-Hist: A Temporal Abstraction of Multivariate Histogram Snapshots

“…Related Work. Over the past decade, histogram snapshot analysis has been performed by applying traditional machine learning algorithms, such as PCA [9], clustering based on histogram distance functions [7], and training random forests for classification [6]. However, these traditional methods do not take into account any changes in the histogram distribution over time, i.e., for each variable of each data entity, there is only one representative histogram.…”

“…On the other hand, there are many cases where histograms are generated and stored over time without a particular analytics task in mind but rather for the purpose of data collection or monitoring; and even without the original data. Examples of such histogram data can be found in the automotive industry and are generated from, e.g., monitoring vehicle compressors [17] or sensors in heavy trucks [6]. One benefit of collecting data in the form of histograms is the ability to store larger amounts of data in sensors and devices with small memory capacity, which is a common requirement and challenge in predictive maintenance.…”