Abstract: For loundness and strong randomness of on-site noise in open spaces of power substations, we proposed a partial discharge acoustic pattern recognition method based on improved denoising autoencoder and DenseNet. Firstly, the characteristic frequency band of the partial discharge sound signal was extracted, and the potential features of the signal by establishing an improved denoising autoencoder were extracted. Secondly, potential feature sequences were transformed into two-dimensional images through the Gramian Angular Field transformation, and a dataset of partial discharge feature maps was established. On this basis, a densely connected network identification model was constructed to perform pattern recognition on the local discharge acoustic signal spectra, achieving accurate identification and diagnosis of partial discharge types under random low signal-to-noise ratio conditions. The partial discharge acoustic signals of four typical defect electrode models were collected by piezoelectric acoustic sensors, and pattern recognition was performed on random partial discharge acoustic signals with a low signal-to-noise ratio. The experimental results show that, compared with the accuracy of methods such as directly constructing recognition models using partial discharge acoustic data and using traditional denoising autoencoders for data dimensionality reduction, the pattern recognition accuracy of this paper is higher, reaching 98.6%.