<p>The nonlinear vibration signals of transformer bushings exhibit high nonlinearity and complexity, making their inherent patterns difficult to effectively analyze using linear methods. Accurately extracting signal features is challenging, which complicates the identification process. To address this issue, a deep confidence identification algorithm based on gradient synchronization analysis is proposed for the nonlinear vibration signals of transformer bushings. The formation mechanism of these signals is complex; by constructing a transformer vibration model, the variation patterns of nonlinear vibration signals in bushings under different operating states can be investigated. Adaptive multi-scale morphological gradient transformation is applied to process the vibration signals of transformer windings, extracting the morphological gradient spectrum and gradient spectral entropy to characterize the nonlinear vibration features, thereby facilitating subsequent identification. The concepts of morphological gradient spectrum and gradient spectral entropy are introduced, with the complexity of the spectrum quantified by computing the morphological gradient entropy of the vibration signal, enabling effective information extraction. The calculated morphological gradient spectral entropy of the transformer winding vibration signal is input into a deep belief network (DBN) model, where deep feature extraction and classification are achieved through multi-layer restricted Boltzmann machine stacking and reverse fine-tuning. By incorporating omnidirectional spectrum technology to optimize the network and enriching feature extraction through multi-channel signal fusion, the accuracy and efficiency of the DBN in identifying nonlinear vibration signals of transformer bushings are further enhanced. Experimental results demonstrate that the proposed algorithm effectively computes the morphological gradient spectral entropy of vibration signals, revealing their variation characteristics and trends. The distortion rate of the vibration waveform is less than 2.5%, and the method effectively identifies nonlinear vibration signals with different peak values.</p>