The rapid advancement of domain adaptation algorithms has significantly accelerated the deployment of intelligent diagnostic technologies. However, existing domain adaptation methods predominantly focus on closed-set fault diagnosis and rarely address data privacy concerns, limiting their applicability in industrial settings. To this end, a universal source-free domain adaptation method is proposed. Initially, a source model is pre-trained using labeled source data. This pre-trained model then processes the target data to decompose the target features into common class components and unknown class components, while simultaneously generating target prototypes and source anchors. Subsequently, the distribution of the unknown class components is estimated using a Gaussian mixture model with two components. Finally, a confidence estimation strategy is developed to derive instance-level decision boundaries by evaluating the distance between target prototypes and source anchors, thereby completing the classification task. Experimental results on gearbox and rolling bearing datasets demonstrate that our approach excels in handling fault diagnosis under varying conditions while ensuring data privacy.