Health diagnosis interprets data streams acquired by smart sensors and makes inferences about health conditions of an engineering system thereby making critical operational decisions. A data stream is a flow of continuous data that face some challenges in data mining. This paper addresses concept drift and concept evolution as two major challenges in the classification of streaming data. Concept drift occurs as a result of data distribution changes. Concept evolution happens when new classes appear in the stream. These changes may cause the degradation of classification results over time. This paper presents an adaptive fusion learning approach to build a robust classification model. The proposed approach consists of three steps: (i) proposed fusion formulation using weighted majority voting (ii) active learning to labels selectively instead of querying for all true labels (iii) distance-based approach to monitoring the movement of data distribution. A diagnosis case study has been used to demonstrate the developed fusion diagnosis methodology.