Multiscale Time-Frequency Sparse Transformer Based on Partly Interpretable Method for Bearing Fault Diagnosis

Table 5

The diagnosis results from different methods.

Working conditions (Nm)	Defect mode	Label	Average accuracy of different methods (%)
			CNN	CNN-LSTM	Bi-LSTM	WRN-16-2	TST	TFT	BAFT	MTFST (proposed)
2	Single defect	A1	86.7	88.3	89.2	90.0	87.5	89.2	90.8	92.5
	Single defect + compound defect	A2	82.5	84.2	88.3	85.0	85.0	87.5	86.7	90.8
5	Single defect	B1	90.0	85.8	87.5	90.0	90.0	89.2	90.8	93.3
	Single defect + compound defect	B2	82.5	79.2	85.0	88.3	89.2	89.2	90.8	90.0
10	Single defect	C1	89.7	90.5	89.6	87.7	90.1	90.5	90.4	92.5
	Single defect + compound defect	C2	83.3	87.5	84.2	87.5	85.8	88.3	89.2	90.8

The bold values indicate the optimal diagnosis accuracy.