Accurate fault pattern recognition under various operating conditions is a huge challenge for vibro-acoustic fault diagnosis of rolling bearings. Traditional feature fusion methods are difficult to extract sensitive features related to faults under various operating conditions. To solve the problem, a novel feature fusion method is proposed in this paper based on entropy-weighted nuisance attribute projection (EWNAP) and orthogonal locality preserving projections (OLPP). Specifically, the method mainly includes three steps: feature extraction, alleviating interference of operating condition, and fusion feature. First, features are extracted from the acquired sound signals. Second, by introducing the covariance matrix of the acquired signals and fuzzy entropy to improve the weighted matrix of nuisance attribute projection (NAP), the EWNAP is proposed to reduce the impact of operating conditions in these extracted features. In the end, OLPP is adopted to fuse features and obtain sensitive features. The clustering performance of the proposed method is quantitatively described by an evaluation index, and for fault pattern recognition, the vectors as samples, which are composed of features extracted via EWNAP-OLPP, are fed into the back propagation (BP) neural network. The analysis on a fault case of rolling bearings shows that the proposed method is robust for vibro-acoustic fault diagnosis of rolling bearings under various operating conditions and superior to traditional methods.