Edwin Lughofer, Patrick Zorn, Edmund Marth,
"Transfer learning of fuzzy classifiers for optimized joint representation of simulated and measured data in anomaly detection of motor phase currents"
, in Applied Soft Computing, Vol. 124, Nummer 109013, Elsevier, 5-2022
Transfer learning of fuzzy classifiers for optimized joint representation of simulated and measured data in anomaly detection of motor phase currents
Sprache des Titels:
The generation of simulation data from physical models trying to mimic parts of the real-world process as accurately as possible has received much attention in industry during the last years. A proper augmentation of simulation data with (available or recordable) real-measured data is an intrinsic challenge to increase the performance of machine learning models. In this paper, we propose a joint representation learning approach based on an optimized transfer of fuzzy classifiers from original simulated data (as source task) to real-measured data (as target task). This is particularly done for binary classification tasks under the scope of anomaly detection of phase currents in electrical machines, where only anomaly-free real-measured data can be collected in advance (so, only in the simulated data anomaly samples are present). Our approach combines distribution matching between simulated and real-measured data on the basis of fuzzy rule activation levels with a weighted error term representing the classification loss between anomaly-free and anomaly data. Both are realized on a local basis per rule, which yields good flexibility during optimization with respect to the actual local distributions and their positions in the target task and is particularly possible for fuzzy classifiers due to the local geometric interpretation of the rules. We demonstrate how to find a feasible initial solution for the numerical optimization process. Results on data sets from the application showed significantly improved classification accuracies of up to 25% compared to several conventional fuzzy classifiers training variants and also compared to various renowned machine learning classifiers, while a respectful (slightly worse) performance compared to the two top performers (SVMs and deep MLPs) were achieved. However, a large performance gap between conventional fuzzy classifiers and these ML-based top performers could nearly be closed and due to the local geometric interpretation of the internal fuzzy partitions, explainable insights into the model transfer could be yielded to experts.