Abstract: To address the lack of oil chromatographic equipment in distribution transformers and the difficulty of applying traditional condition assessment methods relying on multi-source measurement data, this paper proposes a multi-level health warning method based on smart meter data. First, 15-minute active and reactive energy data are used to compute transformer losses, and missing data are reconstructed via Soft-Impute to obtain a continuous loss series. K-means clustering is then employed to decouple temporal features, enabling segmentation of intraday operating patterns. On this basis, an improved Block Studentized Bootstrap method is introduced, combining clustering-based stratification with time-block resampling to preserve temporal dependence and state distribution characteristics. Meanwhile, Studentized normalization reduces variance bias across subsamples under different load levels, improving estimation stability. A segmented loss-based health profile is constructed, from which 15 minute, daily, and monthly warning indicators are derived. Finally, validation using data from four 10 kV transformers in a Northwest China grid and a fault-month case shows that the proposed method can stably trigger alarms during degradation periods, achieving 95.4% precision and an F1-score of 0.91. The method relies solely on smart meter data and demonstrates effective early detection of slowly evolving latent faults.