Abstract: The measurement of liquid hydrogen/liquid oxygen propellant levels in rocket low-temperature is a crucial step in space launch missions.Segmented capacitive sensor is widely used in rocket tank liquid level monitoring due to their simple structure,high precision and reliability.However,due to the complex physical phenomena such as evaporation,boiling,sloshing and thermal stratification during the filling and storage of cryogenic propellants,the traditional section identification method of capacitive sensors based on voltage fitting slope and fixed threshold has poor noise resistance,serious error accumulation and requires manual intervention,which is difficult to meet the requirements of high-reliability autonomous measurement for multiple tanks.To address these challenges,a deep learning method for section identification of sectioned capacitive sensors based on multi-scale feature fusion is proposed.A dual-channel noise suppression module is adopted to achieve high-fidelity denoising and downsampling of the original waveform through cascaded convolution and mixed pooling operations;a residual feature extraction network with channel-space hybrid attention mechanism is constructed to effectively capture the multi-scale dependencies of the time series signal and adaptively enhance the expression of key features;based on the adaptive multi-scale feature fusion module,local details,medium-range correlations and long-range dependencies are integrated through parallel multi-branch convolution and feature concatenation-compression strategies to form a feature representation that is easy to identify.In addition,a data generation framework for the waveform characteristics of sectioned capacitive sensors is established.Through physical process modeling,multi-channel voltage mapping and multi-dimensional noise injection,a large-scale simulation dataset is constructed for model training,and the trained model is verified using real sensor data under actual working conditions.The results show that in the section identification task,the model achieves an average accuracy of 91.94%,and the average height calculation accuracy rate is 92.49%.The model has significant advantages in identification accuracy and noise robustness,and has the ability to correct the number of sections after an incorrect identification.This method provides an effective technical approach for high-precision and high-reliability autonomous section identification of sectioned capacitive sensors under complex working conditions.