Mechanical Fault Diagnostics by History-Status Ensemble Neural Network : Method and its Validation for Bearings

Neural network architecture is crucial for deep learning. Plenty of complicated neural network architectures have been proposed for higher accuracy in the fault diagnosis filed. However, the performance of a carefully designed neural network may decline when applied to the tasks different from that they had been originally intended for. A novel History-Status Ensemble (HSE) method is proposed in this paper . Benefitting from the back-propagation gradient descent algorithm, the proposed method is versatile and can be easily applied to various neural networks without a re-design or tweaking of network architectures. Comparative experiments based on the Convolutional Neural Network (CNN) are implemented, and experimental results show that the proposed method significantly improves the classification accuracy. Besides, the HSE method demonstrates an excellent generalization ability and adaptability to different network architectures as well as robustness with respect to training parameters