A method that integrates the motif difference field (MDF) with an improved EfficientNet was proposed for gearbox fault diagnosis to address the difficulty of diagnosing gearbox faults under complex operating conditions. Firstly, the original fault feature signal was transformed from one-dimensional data into a two-dimensional feature image by using the motif difference field. Then, the two-dimensional feature image was input into a neural network that was based on the EfficientNet framework. This network was further optimized by incorporating the coordinate attention (CA) mechanism, channel concatenation operations, channel shuffle operations, and group convolutions to enhance its feature extraction and expression capabilities, thereby improving the accuracy of gearbox fault diagnosis. In line with the experimental results, the proposed method is shown to be suitable for diagnosing gearbox faults under complex operating conditions, with a fault recognition accuracy being achieved of up to 98.9%. Moreover, the recognition accuracy can still be reached of over 90% even when Gaussian white noise with a signal-to-noise ratio of -10 dB is added.