Fiber optic sensors have been extensively researched and widely applied because of their high sensitivity and immunity to electromagnetic interference. However, in the process of collecting fault acoustic signals from transformers, there is a drawback of relatively high noise content, making signal feature extraction difficult or sometimes impossible. To address this issue, this paper proposes to use information entropy theory for feature extraction analysis of transformer fault acoustic signals and conduct recognition research on the extracted features based on support vector data description (SVDD). An experimental platform for collecting transformer fault acoustic signals is built in the laboratory, and signals from three typical discharge models are acquired. Based on information entropy theory, four types of entropy values, including fuzzy, energy, singular spectrum and power spectrum entropy, are selected to extract features from the filtered acoustic signals, forming recognition feature vectors. Finally, the SVDD algorithm is employed to study the recognition of the extracted features. The results demonstrate that the recognition accuracy of the extracted fault acoustic signal features exceeds 90%, surpassing traditional time-frequency domain feature extraction and wavelet transform-based feature extraction methods. This confirms the feasibility of the proposed approach.