I.3 Related work 35 data analytics. Appendix B.1, Table B.1 (divided into two parts), summarises and compares relevant literature on data-driven methods for the automatic inference of FTmodels. Early techniques for data-driven FT inference include the IFT algorithm (Madden and Nolan, 1994), which employs Quinlan’s ID3 algorithm to generate Decision Trees, and the approach in Mukherjee and Chakraborty, 2007, which uses text mining techniques with maintenance records as input data. Mukherjee and Chakraborty, 2007 addresses the challenge of inferring FTs throughlinguistic analysis anddomain knowledge to extract failure characteristics from brief descriptions of equipment faults. Roth, Wolf, and Lindemann, 2015 propose a method using the Structural Complexity Management methodology to deduce dependencies, which are later used to infer the Boolean logic operators of FT models. Inspired by Causal Decision Trees, the LIFT algorithm (Nauta, Bucur, and Stoelinga, 2018) utilises the MantelHaenszel test to identify dependencies between events, requiring both basic event data and intermediate event failure information. The ILTA (Waghen and M.-S. Ouali, 2019) and MILTA (Waghen and M.-S. Ouali, 2021) algorithms combine Knowledge Discovery in Datasets (KDD), Interpretable Logic Tree Analysis, and Bayesian probability rules. Another method, described in Linard, Bueno, Bucur, et al., 2020, constructs a Bayesian Network before converting it into an FT model, employing blacklists and whitelists to identify the presence or absence of arcs. The DDFTA algorithm (Lazarova-Molnar, Niloofar, and Barta, 2020) derives FTs from failure data time series through binarisation and Boolean equation simplification. The DDFTAe algorithm (Niloofar and Lazarova-Molnar, 2021), an extension of DDFTA, addresses missing information in time series fault occurrence data, while the DDFTAnb algorithm (Niloofar and Lazarova-Molnar, 2023b) extends DDFTA by focusing on FT models using naïve Bayes classification and time series data. Evolutionary algorithm-based methods include FT-EA (Linard, Bucur, and Stoelinga, 2019), FT-MOEA (Jimenez-Roa, Heskes, Tinga, et al., 2023), and FT-MOEA-CM (Jimenez-Roa, Rusnac, Volk, et al., 2024), with FT-MOEA demonstrating improvements through a multi-objective cost function over the singledimensional cost function used in FT-EA, and FT-MOEA-CM proved larger scalability by considering metrics computed from the Confusion Matrix. The SymLearn tool chain (Jimenez-Roa, Volk, and Stoelinga, 2022) enhances scalability by exploiting symmetries and modules within failure datasets. Additionally, Dorfhuber, Eisentraut, and K&etínsk’, 2023 employs genetic algorithms to learn attack trees from sets of traces. The method in Verkuil, Budde, and Bucur, 2022 generates FT models from sensor time series data, exemplified by a domestic heater case study, which aims to identify thresholds that di!erentiate between normal and error conditions. The ITCAmethodology (Waghen and M. Ouali, 2022) focuses on fault hierarchy causality
RkJQdWJsaXNoZXIy MjY0ODMw