Considering that only a small amount of bearing training data can be collected in real-world situations, making it difficult to build a reliable and robust deep learning-based model, and that early fault signatures are often subtle and unobtrusive, an ARMA-based online Early Fault Detection (EFD) method is proposed to solve these two problems. Firstly, an sensitive dedicated health indicator is first found to make early failures apparent. And then, Adaptive Piecewise Constant Approximate Segmentation (APCAS) is introduced to enable the classification of health stages. Finally, experiments are carried out on the IEEE Prognostics and Health Management (PHM) Challenge 2012 bearing dataset. The results show that the proposed method is effective in accurately detecting early faults.