Abstract
In the traditional fault diagnosis method based on convolutional neural network, the dimension of the higher-order input tensor of the pooling layer is reduced under variable working conditions, the tensor data are easily destroyed to cause the loss of data information. In addition, the diagnosis of single modal features will ignore the coupling of fault information under variable working conditions and lack the joint extraction of other modes, so that the model performance is
restricted. To overcome these deficiencies, combining the advantages of the tensor projection layer and multimodality, a new fault diagnosis method based on a multimodal-deep tensor projection network is proposed under variable working conditions. In the proposed method, the multimodal features obtained by modulating and demodulating vibration signals are transformed into a time–frequency map, and the obtained time–frequency maps are fused to construct a third-order tensor composed of time, frequency, and modal number. Then a multimodal-deep tensor projection network is constructed by tensor projection layers instead of pooling layers in traditional deep convolution neural networks. The proposed method avoids the destruction of higher-order input tensor dimension reduction and the loss of information. The recognition accuracy has
greatly improved. The proposed method is verified by the bearing fault diagnosis experiments of speed-up and speed-down processes under variable working conditions, and the inter-shaft bearing fault dataset from an aero-engine system. The experimental results show that the proposed method is very effective. The proposed method contains more dimensional feature information, can better extract fault features and improve the recognition rate of different types of
faults.
restricted. To overcome these deficiencies, combining the advantages of the tensor projection layer and multimodality, a new fault diagnosis method based on a multimodal-deep tensor projection network is proposed under variable working conditions. In the proposed method, the multimodal features obtained by modulating and demodulating vibration signals are transformed into a time–frequency map, and the obtained time–frequency maps are fused to construct a third-order tensor composed of time, frequency, and modal number. Then a multimodal-deep tensor projection network is constructed by tensor projection layers instead of pooling layers in traditional deep convolution neural networks. The proposed method avoids the destruction of higher-order input tensor dimension reduction and the loss of information. The recognition accuracy has
greatly improved. The proposed method is verified by the bearing fault diagnosis experiments of speed-up and speed-down processes under variable working conditions, and the inter-shaft bearing fault dataset from an aero-engine system. The experimental results show that the proposed method is very effective. The proposed method contains more dimensional feature information, can better extract fault features and improve the recognition rate of different types of
faults.
Original language | English |
---|---|
Article number | 112336 |
Number of pages | 24 |
Journal | Mechanical Systems and Signal Processing |
Volume | 225 |
Early online date | 13 Jan 2025 |
DOIs | |
Publication status | Published - 15 Feb 2025 |