A lightweight main circulating pump fault diagnosis model based on multidimensional vibration feature graph is proposed to address the difficulties in feature extraction and the large scale of fault diagnosis models in the cooling system of a ultra high voltage converter. First, a time-domain feature extraction method based on vibration locus images (VLI) and pseudo-color coding is proposed to construct the time-domain feature graph of the main pump. Secondly, by integrating Markov transition field (MTF) and wavelet packet transform (WPT), the low frequency and high frequency fault characteristics of vibration signals are extracted at full scale, and the frequency domain and time-frequency domain feature maps of the main pump are constructed. Finally, a lightweight convolutional neural network model framework is improved through omni-dimensional dynamic convolution, and a lightweight main pump fault diagnosis model (OD-ShuffleNet) is constructed. The model integrates time-domain, frequency-domain, and time-frequency domain fault features, further improving the fault diagnosis accuracy while reducing hardware resource consumption. The results show that the diagnostic accuracy of the model is 95.0%, which is better than that achieved by classical convolutional neural network architectures.