There are problems of easy omission of extracted feature information

low diagnostic accuracy and large computation volume in the fault diagnosis process for the modular multilevel converter (MMC) in power systems. Thus a discretized MMC open-circuit fault diagnosis method with modal time-frequency diagrams and Resnet-BiGRU model is proposed. From the open-circuit fault characteristics

the output phase currents and bridge arm voltages are selected as fault parameters. The improved gray wolf optimization algorithm is used to search the optimal parameters in the process of complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN). Combined with the optimal parameters of CEEMDAN

the fault parameters are decomposed into sensitive and high-quality intrinsic mode function (IMF) components and reconstructed. To fully exploit the sensitive components in the reconstructed signals

those signals are transformed into modal time-frequency diagrams using continuous wavelet transforms; the modal time-frequency diagrams in different fault categories are input to the Resnet-BiGRU model for training

testing and outputting the diagnostic results

so as to complete the diagnosis of the faulty bridge arms and the localization of the faulty IGBT in the sub modules. The experimental results show that: its diagnosis of faulty bridge arms and localization of faulty IGBT in the sub modules reach an accuracy of 98.63% and 99.87%

with high diagnostic accuracy; the diagnostic process possesses a response time of seconds; compared with other methods

the proposed method has a higher accuracy in extreme conditions such as small samples

data imbalance

and noise interference. The work provides a new way for the diagnosis of power system faults.