In the passive recognition task of ship radiated noise, the efficiency of extracting relevant classification information from the noise directly affects the classification results. This study proposed a multidimensional attentional convolution (MAConv) module as a fundamental component for capturing decisive classification information. Drawing inspiration from dynamic convolutional methods such as CondConv, we perform multidimensional attentional integration in three dimensions: input channel, output channel, and spatial channel, which involves a linear combination of different static convolutional kernels. The weights of the combination are data-dependent, with different samples using different weights, so the network will focus on different input channels, output channels and spatial channels depending on the samples in classification. As a drop-in replacement of traditional convolutions, MAConv can be easily integrated into popular CNN architectures, and subsequent experimental results on the open-source dataset Deepship show that CNN networks using the MAConv plugin can significantly improve the performance of ship radiation noise classification.