Classifying Abnormal Behaviour Caused by Soft-errors in Logs From PID-controlled Environments Using Machine Learning Models - A Study on the "LANSCE 18 Cruise" Dataset
Authors
Bonde, Anna Veibel ; Mikkelsen, Simon Kanne
Term
4. term
Education
Publication year
2023
Abstract
Soft errors are transient bit flips, often caused by cosmic radiation, and are becoming more concerning as chip density increases. This thesis investigates whether sequence-based machine learning models can classify soft-error occurrences in logs from a PID-controlled cruise control environment, using the LANSCE 18 Cruise dataset and two related extensions where embedded systems were exposed to an accelerated neutron beam. We pose a binary classification task—detecting whether a soft error occurred in a given log—and compare RNN, LSTM, GRU, ROCKET, and HIVE-COTE 2.0 on both raw time-series data and data transformed using matrix profiles. On the original representation, ROCKET achieved 0.67 accuracy, while transforming to matrix profiles generally improved performance substantially: ROCKET reached 0.96 and GRU 0.99. These findings suggest that machine learning is a promising approach for detecting soft errors in PID-controlled environments and that matrix profiles add valuable information that markedly improves classification.
Bløde fejl (soft-errors) er forbigående bitflips, ofte udløst af kosmisk stråling, som bliver stadig mere relevante i takt med øget chip-tæthed. Denne afhandling undersøger, om sekvensbaserede maskinlæringsmodeller kan klassificere forekomster af bløde fejl i logfiler fra et PID-styret cruise control-miljø, baseret på LANSCE 18 Cruise-datasættet og to tilknyttede udvidelser, hvor indlejrede systemer blev udsat for en accelereret neutronstråle. Vi formulerer problemet som binær klassifikation af, om en blød fejl er opstået i en given log, og sammenligner RNN, LSTM, GRU, ROCKET og HIVE-COTE 2.0 på både rå tidsrækker og data transformeret til matrixprofiler. På den oprindelige repræsentation opnåede ROCKET en nøjagtighed på 0,67, mens transformation til matrixprofiler generelt forbedrede resultaterne markant: ROCKET nåede 0,96 og GRU 0,99. Resultaterne indikerer, at maskinlæring er lovende til at opdage bløde fejl i PID-styrede miljøer, og at matrixprofiler tilfører information, som væsentligt forbedrer klassifikationen.
[This apstract has been generated with the help of AI directly from the project full text]