<p>Early detection of bearing faults is vital for ensuring the reliability and safety of rotating machines such as automobile wheels and power plant turbines, because undetected degradation may lead to unexpected failures and operational risks. As bearings deteriorate, vibration signals exhibit characteristic changes reflecting underlying mechanical issues. Machine learning (ML) and deep learning (DL) have advanced fault detection but often rely on task-specific architectures with extensive labeled datasets and frequent retraining. Recently, large language models (LLMs) have emerged as a complementary approach, offering strong generalization through prompt-based learning. With well-designed prompts and structured inputs, LLMs can perform time-series reasoning and support diagnosis via few-shot or zero-shot learning without retraining. However, limited input capacity of LLMs challenges the handling of long-duration vibration sequences essential for modeling gradual progression. To address this, we propose a four-stage framework enabling compact representation of vibration data and efficient delivery of essential information for LLM-based analysis. First, statistical and spectral features are extracted and compressed using Principal Component Analysis (PCA) to reduce token count while preserving key information. Second, data is reformatted into natural-language-compatible input while maintaining temporal structure. Third, a Multi-Strategy Retrieval-Augmented Generation (Multi-Strategy RAG) architecture processes large-scale time-series data. Finally, task-specific prompts guide the LLM to reasoning tasks such as detection and interpretation. The proposed input data strategy leads to enhanced fault diagnosis performance across publicly available bearing benchmark datasets.</p>