Deep forest based multivariate classification for diagnostic health monitoring

Diagnostic health monitoring without prior knowledge is still a hard problem in the Prognostic and Health Management (PHM) field. A traditional approach is to construct one-dimensional Synthesized Health Index (SHI) and set the thresholds of different system health stages referred to empirical values. However, there is still a difficult problem to be solved on choosing appropriate SHI and thresholds for a complicated system, which are the key factors seriously influencing the result of diagnostic health monitoring. Compared with this approach, a new method of deep forest is proposed to achieve multivariate classification for diagnostic health monitoring. Firstly, an unsupervised feature selection method based on spearman's correlation is improved to remove redundant features automatically. Then, the k-means algorithm is introduced to acquire stage knowledge of degradation process, therefore achieving supervised learning. Finally, deep forest method is adopted to train the data samples to obtain the classification model of system health for online diagnostic health monitoring. The verification results on NASA data sets demonstrate that the proposed diagnostic health monitoring scheme is effective and feasible.