Tensor Poincaré plot index: A novel nonlinear dynamic method for extracting abnormal state information of pumped storage units

Abstract

Efficiently extracting information from the massive data that characterize the abnormal condition is an important topic for pumped storage units (PSUs) operation and maintenance. Existing feature extraction methods for PSUs have weakened the connection between time and frequency domain features of signals, and the extracted information cannot fully represent the PSU operational state. Therefore, the paper proposes tensor Poincaré plot index (TPPI), a feature extraction method for quantifying PSU operation on multiple time and frequency scales. Firstly, the operational datasets are hierarchically decomposed and coarsely granulated to obtain components at different time and frequency scales. Secondly, the different components are sequentially transformed into Poincaré plots, and the key indexes of these plots are extracted, respectively. Finally, the proposed model is constructed by the extracted features and random forests. The proposed model is applied to two case of hydraulic anomaly identification and mechanical fault diagnosis, based on the measurement of the actual PSUs. The results show that indicators of this method are no less than 99.629 % and 99.660 %. In comparison experiments with 15 popular methods, the proposed model exhibits superior competitiveness, robustly affirming the advantages of the TPPI. The proposed method is helpful for promoting the intelligent construction of PSUs.