2023-04-19
2023-04-19
Tianfu Li Research
Feature Fusion
(wangRollingBearingFault2022, link, DOI, zolib)
(liConvolutionalNeuralNetworkbased2022, link, DOI, zolib)
Feature-fusion: After two convolutional and pooling layers, the feature maps acquired from multiple sensors are then combined and fed into the next two convolutional and pooling layers. Finally, a fully connected layer and a linear layer are performed to generate the image classification results.
Decision-fusion: Since the output of each CNN model is the probability of a certain type of quality level, in this work, a simple decision fusion method is developed by calculating the maximum probability to make the final conclusion about the quality monitoring result.
(zhangMLPCCNNMultisensorVibration2022, link, DOI, zolib)f
(zhangBrakeUnevenWear2022, link, DOI, zolib)
Ensemble depends on the assumption that the uneven wear condition of friction block in the braking system induces multi-source vibration responses in the braking process of high-speed trains. The prediction results of all models can be combined to obtain a more accurate description of the partial wear state, and the most direct integration method is the weighted average of each prediction result.
A grid search algorithm reference was employed to optimise the parameters and find an optimal set of weight combinations α1′, α2′ and α3′ from the validation set data
(guMotorOnLineFault2023, link, DOI, zolib)
