Traditional principal component analysis

relative principal component analysis and other multivariate statistical methods based on threshold to do the fault diagnosis. Since multivariate statistical method is an equivalent representation of the original space

it does not add any amount of information

making it difficult to diagnose minor faults. And the original space is reduced dimensionally into the principal component space and the residual space

so that the tiny information cannot be fully expressed. Deep learning has been successfully applied in pattern recognition. However

multilevel networks of deep learning represent linear combinations of details but do not have explanatory. Only with the result of training without theoretical basis

it is difficult to carry out mechanistic analysis. This paper presents a fault diagnosis method which combines PCA thought and deep learning thought. Based on the original PCA

this paper first expands and then reduces the dimension

making the inexplicit information in the original space fully expressed and interpreted. Theoretical and simulation experiments show that this method can judge the minor faults which cannot be detected by traditional PCA

improve the detection rate of fault detection and have interpretability.