Log anomaly detection plays a critical role in maintaining the stability and security of information technology systems. Current methods often rely on single-model architectures, limiting accuracy and generalization, especially with complex and diverse log data. Furthermore, these approaches typically lack effective feedback mechanisms for continual performance enhancement, which hampers their scalability in practical applications. To address these challenges, we introduce LogMoE, a novel model that integrates autoregressive models with mixture-of-experts networks. This model aims to enhance adaptability to diverse log data and improve anomaly detection precision. The autoregressive component captures temporal dependencies within log sequences, while the mixture-of-experts network effectively handles diverse log patterns. To optimize computational efficiency, we employ knowledge distillation techniques for model parameter compression. Additionally, human feedback reinforcement learning is introduced to refine the feedback mechanism, further enhancing model performance and scalability. Extensive experiments conducted on multiple public datasets validate the efficacy of LogMoE. Results demonstrate significant improvements in detection accuracy and generalization capability compared to existing methods.