Speaker
Mr
Robert Szmurło
(Warsaw University of Technology)
Description
We propose a method for improving classification for faults detection in electrical machines when the training data contains a significant amount of wrongly labeled inputs. The proposed method uses the ROCKET - fast and accurate time series classification using random convolutional kernels model logits as an input for classical classification models (neural networks, SVC). The logits from ROCKET model are used for improving the labelling. We study the accuracy of fault detection with different level of wrong labelling. The paper analysis binary classification for fault detection and also multiclass for a few kinds of faults (mechanical and electrical). The input data is collected from publicly available dataset.
Authors
Mr
Robert Szmurło
(Warsaw University of Technology)
Stanisław Osowski
(Warsaw University of Technology)
