Abstract The sensitivity of various features that are characteristics of machine health may vary considerably under different operation conditions. Thus it is critical to devise a systematic feature selection (FS) scheme that provides a useful and automatic guidance on choosing the most representative features for machine health assessment without human intervention. This paper presents a hybrid feature selection scheme named HFS based on the combination of Gaussian mixture models (GMM) and K-means. The proposed scheme is based on unsupervised learning technique, which does not need too much prior knowledge to improve its utility in real-world applications. The effectiveness of the scheme was evaluated experimentally on bearing test-beds, using degradation prediction and visualization approach where self-organizing map (SOM) model was employed. A novel health assessment indication, i.e., log likelihood probability (LLP) is proposed to provide a comprehensible indication to quantify machine health states. A Bayesian-inference-based failure probability (BIP) approach is proposed to calculate the failure risk probability of machines. This paper focuses on identifying the health states of the key machine component (i.e., bearing) under the assumption that the predictable abnormal patterns are not available. The proposed scheme has shown to provide the good machine health assessment performance with reduced feature inputs. The experimental results indicate its potential applications as an effective tool for machine health assessment.