Aiming at the problems of low Test Case Acceptance Rate （TCAR） and lack of diversity in application of fuzzing in Industrial Control Protocols （ICPs）， a fuzzing method for ICPs based on adaptive dynamic interval strategy was proposed. Recurrent Neural Network （RNN） was added to self-attention mechanism in Transformer to construct a protocol feature extraction model； RNN was used to extract local features of the data through a sliding window， and the self-attention mechanism was introduced to carry out global feature extraction， so as to ensure the TCAR； the residual connection was added between the attention blocks to transfer the weight scores and improve the computational efficiency； a dynamic interval strategy was generated to adjust sampling range of the model at any time step， so as to increase diversity of the test cases； in the testing process， the field adaptive importance function was constructed to locate the key variant fields. Based on the above method， a fuzzing framework TDRFuzzer was designed and experimentally evaluated using three industrial protocols： Modbus TCP， S7 comm， and Ethernet/IP. The results show that compared to three models： GANFuzzer， WGANFuzzer， and PeachFuzzer， TDRFuzzer has the TCAR increased significantly， and the Vulnerability Detection Rate （VDR） increased by 0.073， 0.035， and 0.150 percentage points， respectively. This indicates that TDRFuzzer has stronger vulnerability mining capability for ICPs.