Aiming at the detection and recognition challenges of underwater unmanned underwater vehicle(UUV) radiation noise, a proposed method combining adaptive feature modal decomposition(AFMD) with a joint convolutional neural network(JCNN) is proposed. Firstly, AFMD is applied to the signal to obtain a set of decomposition components, followed by reconstruction based on the GiNi index to select the most optimal component. Then, the reconstructed signal undergoes continuous wavelet transformation, producing 2D time-frequency maps representative of distinct radiation noise categories. The methodology incorporates a feature fusion module within the frequency dynamic convolution module and SGE module to establish the JCNN. This network adeptly extracts features from the 2D time-frequency maps, facilitating the classification of UUV radiation noise. Experimental results demonstrate the effectiveness of the proposed method in accurately identifying underwater UUV radiation noise, achieving a notably high recognition rate.