To address the issues of structural complexity, variable working conditions, frequent failures, and the presence of noise and irrelevant information in the monitoring data of a certain type of diesel engine, this paper proposes a signal processing denoising method based on Dung Beetle Optimization (DBO) and Variational Mode Decomposition (VMD), combined with an Interpretable Random Forest (IRF) to construct a fault diagnosis model. In this method, DBO is utilized to optimize the decomposition layer number and penalty parameter of VMD. The signal is reconstructed based on the kurtosis criterion of the decomposed modes, and the IRF model is employed to diagnose diesel engine faults. Comparative experiments were conducted with six other fault diagnosis models, such as EEMD-IRF and PSO-VMD-SVM. Results show that the proposed DBO-VMD-IRF method achieves an average accuracy of 99.18% and demonstrates significant advantages in diagnostic performance, accuracy, and reliability. This study provides reliable technical support for fault diagnosis of armored diesel engines.