Detecting errors in the ATLAS TDAQ system: A neural networks and support vector machines approach

摘要

This paper describes how neural networks and support vector machines can be used to detect errors in a large scale distributed system, specifically the ATLAS Trigger and Data AcQuisition (TDAQ) system. By collecting, analysing and preprocessing some of the data available in the system it is possible to recognize and/or predict error situations arising in the system. This can be done without detailed knowledge of the system, nor of the data available. Hence the presented methods could be used in similar system without significant changes. The TDAQ system, and in particular the main components related to this work, is described together with the test setup used. We simulate a number of error situations in the system and simultaneously gather both performance measures and error messages from the system. The data are then preprocessed and neural networks and support vector machines are applied to try to detect the error situations, achieving classification accuracy ranging from 88% to 100% for the neural networks and 90.8% to a 100% for the support vector machines approach.