Automated Anomaly Detection in Data Streams on Engine Test Benches Using Multi-Scale Convolutional Recurrent Encoder-Decoder

Authors

J. Juhl MSc, T. Rudolf MSc, L. Kohout MSc, Dr.-Ing. S. Schwab, FZI Research Center for Information Technology, Karlsruhe; Dr.-Ing. G. Dummer, IAVF Antriebstechnik GmbH, Karlsruhe; Prof. Dr.-Ing. S. Hohmann, Prof. Dr.-Ing. W. Stork, Karlsruher Institut of Technology (KIT), Karlsruhe:

Year

2021

Print Info

Production/Publication ÖVK

Summary

The analysis of engine test bed measurements requires the evaluation of a large number of measured signals as data streams. The data streams from the device under test pose an increasing challenge for test engineers to handle the amount of data available. An evaluation regarding anomalies is only possible in real time if the majority of the test system is automated. Automated methods often use statistics or semi-statistical feature-based methods like signal means, variances, or bounds. However, the performance of these methods is limited on highly dynamic data. This article proposes an architecture to automatically evaluate aggregated data for anomalies in combustion engine test bench system. For this purpose, the artificial intelligence method of the multi-scale convolutional recurrent encoder-decoder with a novel adaptation of the neural Network for additional meta data is used to detect anomalies in the measurements as well as identify the causal signals. Thus, the challenging task for a test engineer can be mode more comfortable by supervising the relevant signals, or a stop event can be triggered automatically in a critical anomaly score. Conclusively, the testing process efficiency when working with multi-signal data streams is increased and damages of the device under test through faults may be mitigated by the early detection serious anomalies.

Number of pages

13 Order Members Login