In order to solve the identification and diagnosis of transformer iron core looseness faulta method of removing environmental noise based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)-wavelet-threshold is proposedand a method of identifying transformer iron core looseness fault based on voiceprint using three-dimensional convolutional neural network (3D-CNN) is proposed. A transformer iron core looseness fault test platform is built to collect the noise signals of the iron core under different degrees of looseness. The voiceprint signal collected for fault identification is filtered by CEEMDAN-wavelet-threshold algorithmand the transformer voiceprint signal with high signal-to-noise ratio is obtained by using the difference between transformer body noise and environmental noise in kurtosis. Thenthe time-frequency matrix is generated by short-time Fourier transform of the filtered voiceprint signalthe Mel spectrogram is obtained by dimensionality reduction of Mel filterand the data set suitable for the input format of 3D-CNN is made. Each layer of the network is builtand 3D-CNN is used to classify and identify the transformer iron core looseness fault. The experimental results show that the recognition rate of transformer iron core looseness fault is more than 90% under the condition of considering environmental noiseand can be used for the recognition and diagnosis of transformer iron core looseness fault.