To address the critical limitations of conventional methods for detecting trace moisture content in transformer oil-such as destructive sampling and poor anti-interference capability-this study proposes a non-invasive, high-precision detection technique based on the integration of broadband ultrasonic time-frequency analysis and deep learning. The approach combines broadband multi-frequency ultrasonic scanning with deep neural network modeling to dynamically characterize and quantitatively predict complex acoustic signatures associated with trace moisture in oil. A dataset comprising 240 transformer oil samples was established, with ground-truth moisture content calibrated using a precision trace moisture detector (220 samples for training, 20 for testing). Each sample was subjected to broadband ultrasonic excitation at six distinct center frequencies. Echo signals were processed via Continuous Wavelet Transform (CWT) for time-frequency analysis, extracting a 128×1 000-dimensional joint high/low-frequency feature matrix as model input, with actual moisture content as output. The core innovation lies in constructing a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) deep learning model: The CNN branch efficiently extracts spatial patterns from the CWT time-frequency spectrograms, while the LSTM branch captures temporal dynamics and cross-frequency dependencies within ultrasonic features. This synergy establishes a robust nonlinear mapping between complex acoustic characteristics and moisture content. Comparative experiments with multiple models demonstrate the superior performance of the CNN-LSTM framework, achieving an exceptionally low mean absolute error (MAE) of 1.33 mg/kg, a mean absolute percentage error (MAPE) of 7.167%, and a high coefficient of determination (R2) of 0.958. This research provides a novel, industrially viable solution for online, non-destructive, and high-accuracy monitoring of trace moisture in transformer oil.