Abstract: As submarine oil and gas are exploited further, the safety of submarine pipelines is receiving increasing attention. Due to the complex operating environment and harsh working environment, submarine pipelines are vulnerable to damage; this leads to accidents. Once an accident occurs in the submarine pipeline, it not only causes massive economic losses but also adversely affects marine ecology. The metal magnetic memory (MMM) technology was proposed in the 20th century to detect macro defects and hidden defects early. To overcome the difficulties of the MMM quantitative inversion of submarine pipeline defects, this study proposed a gated recurrent unit (GRU) neural network model based on improved particle swarm optimization (IPSO). The X52 pipe specimens with blind plates that were welded at both ends were used, pipes had prefabricated defects of different diameters and depths. An 11-6W noncontact probe was used for underwater testing; the host was the TSC-5M-32 MMM Instrument. After conducting simulated submarine tests to obtain the MMM signals of pipe defects, the characteristic parameters of MMM signals with different defect sizes were extracted. It is found that the MMM characteristic parameters exhibit a complex nonlinear variation for different defect dimensions. Exploiting the GRU’s dual-gate structure that can remember the signal characteristics of defects and its superior nonlinear regression fitting ability, a quantitative MMM GRU inversion model was established for detecting submarine pipeline defects. Furthermore, considering the randomness of the hyper-parameter selection in the model, the IPSO algorithm was used to optimize the hyper-parameters. Validation results show that the model has an average accuracy of up to 96% and 93% for defect depth inversion and defect diameter inversion, respectively. Using the MMM method, this study provides a new idea and method for the quantitative identification and defect inversion of submarine pipeline defects.