Dynamic regularization discriminant local preserving projection method for fault diagnosis

Due to the high dimensionality, serial correlation, and nonlinearity of industrial process data, the primary task for diagnosing fault is to extract key fault features from fault datasets. In this paper, to obtain much more inherent fault information, a dynamic regularization discriminant local preserving projection approach (DRDLPP) based on feature reduction is put forward to diagnose fault, which addresses the small sample size problem of discriminant local preserving projection (DLPP) by incorporating the regularization term into the objective function of DLPP. The enhanced performance of DRDLPP for fault diagnosis over conventional diagnostic approaches mostly benefits from two aspects: One aspect is that DRDLPP can discover local manifold fault information hidden in original sample space by preserving the local neigborhood structure of data; The other aspect is that DRDLPP has the remarkable capacity to capture dynamic information by extending the observation vector with previous observation vectors. What is more, the information criterion function is utilized to capture the optimal dimensionality reduction order and time lag of DRDLPP method. The experimental results of the Tennessee Eastman process demonstrate that the proposed DRDLPP approach provides a better visual performance and achieve lower misclassification rates in fault diagnosis.