Addressing challenges in the hydraulic systems of reverse-pitch drilling rigs—such as the difficulty in early motor fault identification, weak signal characteristics, and complex operating conditions—this paper proposes a Multi-Attention Multi-Modal Convolutional Neural Network (MAMCNN) motor fault diagnosis method. This approach centres on monitoring the current signal of the hydraulic pump drive motor. It employs sliding window segmentation to extract raw time-domain waveforms, frequency-domain energy spectra (FFT), and multi-dimensional statistics. Continuous wavelet transform (CWT) is then applied to convert the signal into a two-dimensional time-frequency image, comprehensively reflecting energy distribution and frequency disturbance characteristics during the fault process. The constructed neural network comprises two parallel branches: the one-dimensional branch models the dynamic evolution of the current signal through a convolutional neural network (1D-CNN), bidirectional gated recurrent units (BiGRU), and an attention mechanism; the two-dimensional branch takes the CWT image as input, extracting time-frequency texture information via a two-dimensional convolutional layer and a channel attention module. Following feature fusion, both streams are fed into a fully connected classifier to achieve high-precision identification of common hydraulic pump motor faults (such as rotor bar breakage and eccentricity faults) versus normal operating conditions. Experimental results demonstrate the method's robust interference resistance and generalisation capabilities, rendering it suitable for intelligent monitoring and early fault warning of reverse-circulation drilling rig motors under complex operating conditions.