Modular multilevel matrix converter （M3C） is a low-frequency power transmission AC-AC converter used for offshore wind power generation. In order to improve the reliability and stability of M3C operation, it is necessary to have an efficient and accurate diagnosis method for the open circuit fault of IGBT in its submodules. Therefore, a deep learning fault diagnosis method based on the combination of convolutional neural network （CNN） and gated loop unit （GRU） is proposed. On the basis of analyzing the operating conditions of M3C submodule, wavelet packet analysis is performed on the original fault data, the high-frequency components of which are converted into two-dimensional fault images through temporal image conversion as the training and validation dataset for deep learning, and the features of the high-dimensional data are extracted by CNN, and then the data is optimized and trained by GRU, so as to realize the diagnosis identification of the M3C fault categories. Compared to traditional methods, this method has more accurate and fast fault diagnosis capabilities.