A Dual Robustness Projection to Latent Structure Method and Its Application

The robust performance of traditional partial least squares (PLS) method is poor because the objective function of PLS is to maximize the square of L2norm between the input and output data spaces, and this method is sensitive to outliers but insensitive to local features and causes nuisance alarms. Conversely, the L1norm can perfectly maintain the signal relative size in global statistical features extraction, and its feature extraction results are robust to outliers. Therefore, in this article a novel dual robustness projection to latent structure method, which includes the robustness of feature extraction and regression coefficients, based on the L1norm (L1-PLS) is proposed. Then, more attention is paid to the robustness of both the input and output spaces and the robustness relationship between those spaces in the dual robustness L1PLS method. The corresponding quality-relevant monitoring strategy is also established. Finally, the robustness of the L1-PLS method is illustrated with the Tennessee Eastman process simulation platform. The results show that the proposed method is insensitive to outliers and maintains the signal relative size, and the fault monitoring results are consistent with the actual situation.