A fault diagnosis method by using extreme learning machine

It is difficult to be directly measured for some product quantities by sensors in industrial processes. There are many ways to use the relationship between process variables and quality variables to predict product quality information indirectly, and then use it to fault diagnosis, such as partial least squares (PLS), total projection to latent structures (T-PLS) algorithm and so on. T-PLS decomposes the principal component space into two subspaces: Y-related subspace and Y-unrelated subspace, according to the prediction value of quality variables based on PLS. This paper presents an improved method of T-PLS. The improved method uses the ELM theory to predict quality, then the projection space is further decomposed based on the quality predict results of ELM. According to the comparison of ELM and PLS as well as the comparison of T-PLS and the new method in this paper, it proves the validity of the proposed method. Simulation verifies its properties.