Industrial control protocols(ICP) have a large number of vulnerabilities due to the lack of authentication, authorization, and encryption measures, which seriously threaten the safety of industrial control systems(ICS). Fuzzing, as a common method for vulnerability mining, has the disadvantages of a low reception rate and insufficient diversity of generated test cases in ICPs. To address these issues and enhance the efficiency of ICPs vulnerability mining, this paper proposed a fuzzing methodology for ICPs based on multi-scale latent feature representation. It combined Transformer and GAN in the latent space. It used the Transformer to obtain vector representations of the latent features of protocol messages, while a dynamic multi-scale discriminator captured semantic information at different scales in the latent representations, fusing local field features and global semantic features, thereby improving the acceptance rate of test cases. Additionally, it introduced a self-adversarial learning strategy to train the GAN, reducing redundancy in latent feature representations, and increasing the diversity of test cases. Based on the above mentioned approaches, it constructed a universal fuzzing framework, MLFRFuzzer, that could be applied to most ICPs, and verified its performance on several typical ICPs, including S7comm, Ethernet/IP, and Modbus/TCP. Experimental results show that MLFRFuzzer achieves higher test case reception rates and greater diversity, and outperforms DCGANFuzzer, WGANFuzzer, and PeachFuzzer by approximately 23.76%, 44.07%, and 71.96%, respectively, in terms of the test target anomaly rate, validating the effectiveness and generalizability of MLFRFuzzer. In addition, it demonstrates stronger vulnerability mining capabilities than traditional ICPs fuzzing methods.