This paper proposes ARDiff, an adaptive reverse-step diffusion framework for unsupervised denoising of vibration signals under high-noise industrial conditions. Unlike conventional diffusion models that rely on fixed denoising trajectories or external condition inputs, ARDiff enables adaptive noise-aware reconstruction without any class labels or guidance, through the integration of three key components. First, the frequency attention module (FAM) introduces a complex-valued attention mechanism in the frequency domain, enabling the model to effectively capture fault-relevant spectral features critical for signal reconstruction. Second, the noise step prediction model (NSPM) estimates the optimal reverse diffusion timestep based on the input noise level, improving reconstruction accuracy. Third, the variance-based scaling method (VSM) aligns the scale of the input signal with the expected distribution of the diffusion model at the predicted timestep, ensuring stable and consistent restoration. Experimental validation using the open-source CWRU dataset, the experimental HAI-PUMP dataset, and the SNU Planetary Gearbox dataset demonstrated that ARDiff consistently outperformed existing denoising models, particularly under extremely low signal-to-noise ratio (SNR) conditions. Ablation studies confirmed the individual contributions of FAM, NSPM, and VSM, verifying that ARDiff is a robust and generalizable denoising framework capable of operating effectively without any external conditions or label supervision. Moreover, ARDiff consistently maintained stable denoising performance under both Gaussian and non-Gaussian noise conditions, highlighting its practical robustness in realistic industrial environments.