Abstract: The web thickness of solid propellant grain is a key parameter for characterizing the combustion process and evaluating burning rate characteristics of solid rocket motors. Aiming at severe combustion interference, low signal-to-noise ratio (SNR), and poor adaptability of traditional fixed-rule algorithms under non-stationary conditions in ultrasonic burning rate tests, an ultrasonic measurement model for propellant grain web thickness was proposed, which integrates adaptive denoising, convolutional neural network (CNN), bidirectional gated recurrent unit (BiGRU) and convolutional block attention module (CBAM). Taking the collected ultrasonic echo signals as model input, the model suppresses combustion-induced noise and retains key echo features via adaptive wavelet threshold denoising. In the network architecture, the model integrates CNN, BiGRU, and CBAM mechanisms to realize joint modeling of local echo features, temporal correlation features of echo sampling sequence, and critical feature responses. Furthermore, a composite loss function based on peak-position reference constraints was introduced to enhance the model characterization capability for effective echo features and improve the prediction stability of grain web thickness under complex working conditions. Experimental results demonstrate that the proposed model achieves a mean absolute error of 0.51 mm and an offset error of 0.46 mm for the grain web thickness measurement on the test dataset. Ablation experiment further validate the effectiveness of each module, indicating that the method exhibits high measurement accuracy and robustness under complex working conditions. It provides reliable technical support for the offline processing of echo data, dynamic inversion of propellant grain web thickness and combustion process analysis in ultrasonic burning rate tests.