Abstract: Aimed at addressing the reliance on manual inspection and the lack of effective trend prediction in oil level monitoring for Air Traffic Control (ATC) radar antenna drive systems, this study proposes a non-contact prediction method that integrates image recognition with a Long Short-Term Memory (LSTM) network. The oil level is digitally captured in real time using the Canny edge detector and Hough transform, enabling automatic identification and quantification. A double-layer LSTM model, incorporating an early stopping strategy to enhance generalization, is constructed to model the oil level time series. Experimental validation based on actual operational data compares the proposed model with ARIMA and single-layer LSTM benchmarks. The results show that the proposed model achieves a Mean Absolute Error (MAE) of 0.0044, a Root Mean Square Error (RMSE) of 0.0072, and a Coefficient of Determination (R²) of 0.9569. Its prediction error is reduced by 56.77% compared to the ARIMA model and by 81.02% compared to the single-layer LSTM model. This research realizes non-contact, high-precision perception and reliable trend prediction for the oil level in ATC radar systems, providing effective technical support for advancing predictive maintenance practices.