Adaptive learning for single-output complex systems via data augmentation and data type identification

Abstract

An adaptive learning approach is proposed for single-output complex systems (SOCS) with two features: data augmentation (A) and data type identification (T). Data augmentation is used to handle small-sized data gathered from multiple phases. Three data types, that is, normal, critical, and noisy data, are identified based on the causal logic relation between the input and single output of complex systems. Moreover, the proposed approach is implemented in an adaptive manner because the testing dataset in the current phase is combined with the training dataset to form a more comprehensive training dataset to predict the output of the next phase. Therefore, the new approach is called adaSOCS-A-T, which first implements data augmentation followed by data type identification. A practical case of building tilt rate (BTR) prediction in metro tunnel construction is studied. The case study results show that the proposed approach can produce more accurate prediction results compared with the other approaches. The proposed adaSOCS-A-T approach is further validated by (1) the comparative superiority of adaSOCS-A-T over adaSOCS with no features, adaSOCS-A with only data augmentation, adaSOCS-T with only data type identification, and adaSOCS-T-A with data type identification followed by data augmentation, (2) comparative results using MIXUP, SMOTE, and Gaussian as the data augmentation methods, and (3) further statistical t-test results.