As the core equipment of high-voltage direct current transmission system, the operation reliability of normal direct current converter valve directly affects the stability of power system. In this paper, a composite processing strategy integrating tensor decomposition theory and deep learning is proposed for the fault diagnosis problem under the missing data scenario of normally straight converter valve. A multidimensional data interpolation model is constructed based on Tucker decomposition, and the efficient recovery of high-dimensional missing data is achieved through the co-optimization of core tensor and factor matrix. Design 1DCNN-BiLSTM hybrid network with attention mechanism to enhance the time-frequency characterization of fault features. As verified by the analysis of simulated and measured vibration data on the PSSE platform, the average relative error of the Tucker decomposition and its rate of change are both minimized in the comparison models, and the average MRE is 2.29 in the random missing data scenario. The MRE is reduced by 38.75% compared to the suboptimal model in the PMU fault scenario with 25% high missing rate. The method in this paper can successfully isolate the fault features of severe faults. Moreover, there are rich fault feature modulation bands in addition to the fault feature frequency.
