To address the issues of premature convergence and insufficient parameter adaptability of the Rime Optimization Algorithm (RIME) when dealing with high-dimensional mixed variables, as well as the problem of the LSTM's reliance on manually set hyperparameters which affect the diagnostic performance, a fault diagnosis method based on the improved RIME algorithm optimizing LSTM is proposed and applied to the multi-condition fault identification of automotive gearbox rolling bearings. Firstly, this study makes three key improvements to the original RIME algorithm to overcome its premature convergence and insufficient parameter adaptability when dealing with high-dimensional mixed variables: introducing an adaptive parameter adjustment strategy based on population diversity to balance the exploration and exploitation capabilities of the algorithm; designing a Levy flight global disturbance mechanism to enhance the ability to escape from local optima; proposing a mixed search space encoding scheme to efficiently handle continuous and discrete hyperparameters. Simulation experiments on standard test functions show that the improved RIME algorithm has significantly improved convergence speed and solution accuracy. Secondly, on this basis, the improved RIME algorithm is applied to optimize the key hyperparameters of LSTM, and an improved RIME-LSTM fault diagnosis model is constructed. Finally, experimental research is conducted using the self-built comprehensive simulation test bench for automotive gearbox rolling bearing faults. The experimental results show that the proposed improved RIME-LSTM fault diagnosis method has a diagnostic accuracy rate of up to 99.98% and an F₁ score of 99.75%, which is significantly superior to the comparison methods such as PSO-LSTM and GWO-LSTM. At the same time, the noise robustness test shows that this method can still maintain the highest diagnostic performance under different noise levels.