The invention discloses a modular five-level current converter fault locating method based on a depth convolution network, and the method comprises the steps: combining capacitor voltage signals collected by submodules of a modular five-level current converter into a multi-channel sequence; carrying out the sampling of the multi-channel sequence, obtaining a data tape and carrying out the normalization processing, wherein the processed data tape is taken as a gray-scale map and serves as the input of a depth convolution network model; extracting the features of data through a plurality of intersected convolution layers and ponding layers in the depth convolution network model, transmitting a characteristic pattern of the last ponding layer to a full-connection layer for fusion, and achieving the fault classification through a softmax classifier, wherein different types are corresponding to fault submodules at different positions, thereby achieving the detection and position of a fault. According to the invention, the data needed by the method is easy to obtain, and there is no need of an additional sensor, thereby greatly reducing the cost. Moreover, the method is stronger in capabilities of MMC fault recognition and location.