Bearing fault detection using deep neural network and weighted ensemble learning for multiple motor phase current sources

The objective of this study is the automatic bearing fault detection of permanent magnet synchronous motors (PMSM) using phase current data. Our research proposes a method using sensor fusion to improve the information quantity as well as the gained quality from all available sensor sources using a multi stage workflow. As an initial feature extraction stage a deep neural network architecture based on a 1D-CNN-LSTM is applied on the raw current data to create baseline probability distributions. Then, probability merging is applied to combine the results of all available baseline classifiers to a new feature matrix which is considered as the feature-set for the final classification stage which is build up on an multi learner ensemble of k-Nearest-Neighbor classifiers. To give all ensemble participants a factor of trust, the ensemble predictions are weighted using either average or an optimized weighting. The proposed method reached accuracies of 98,93% on a public available open source benchmark bearing fault dataset. The main contributions of our research are: 1) Tackling the problem of automatic bearing fault detection for PMSMs based on multiple phase current data 2) Improving the final classification results through optimized weighting giving each classifier a factor of trust 3) Reducing the cross working condition accuracy loss by means of an ensemble reaching a higher level of generalization between different domains.