Predictive maintenance and condition monitoring for remote heavy machinery are compelling endeavors to reduce maintenance cost and increase availability. Beneficial factors for such endeavors include the degree of interconnectedness, availability of low cost sensors, and advances in predictive analytics. This work presents a condition monitoring platform built entirely from open-source software. A real world industry example for an escalator use case from a large railroad company underlines the advantages of this approach. In particular, it is shown that even in situations with initially scarce amounts of data accurate predictions can be made using a hybrid analytics approach. Therefore, it combines neural network training and random forest classifier training using two different data sources. This caters for fast time-to-market and highly accurate predictions.
Deutsche Bahn operates a large number of machines of both, rolling stock and landside infrastructure. Machines finally break and require maintenance, repair, or replacement. The traditional approach of periodic maintenance cycles is increasingly replaced by demand-oriented maintenance, as it turns out to decrease maintenance cost and simultaneously improves on the availability. Demand-oriented maintenance clearly requires a notification system, which continuously monitors a machine's condition, detects failures or deviations from the normal machine state, and informs maintenance personnel. While a vast amount of sensors that can do the monitoring job exist, our focus was directed on a universal, non-intrusive, commodity sensor technology. Consequently, we focus on acoustic emissions and sense those using microphones. These emissions are analyzed and classified using state of the art machine learning technology, neural networks foremost. The advantages of a machine learning approach include the generalization of classification models, such that not every single machine requires individual fitting. Transitioning from a proof-of-concept phase, saving potential and our customers request a short time to market for which convolutional neural networks (CNNs) are not particularly well suited because the availability of labelled data is still poor. The need to develop methods to cope with such situations arises. Hence, we have picked up the notion of transfer learning and show the superiority of this approach in regards of accuracy and required data amounts opposed to learning CNNs from scratch. In order to do so, for both data science and production systems, we utilize a comprehensive stack of open source software technologies that support us in data ingestion, edge computing, data preprocessing, neural network training, software deployment, and visualization tasks.
Speakers: Daniel Germanus Felix Bert