In most approaches for journal bearing rotor system diagnosis, dominant features are manually extracted based on expert’s experience and domain knowledge. With the adoption of advanced journal bearings and the limited knowledge about physics-of-failure, the current practice for feature extraction showed limitations for real applications in the power plant industry. To this end, this paper proposes an unsupervised scheme to extract features from correlated vibration signals. First, raw vibration signals from a pair of sensors are preprocessed by generating two-dimensional images. Second, the vibration images are characterized with a HOG (histogram of original gradients) descriptor. Then, deep learning is used to extract relevant features for journal bearing rotor system diagnosis. To demonstrate the validity of the proposed unsupervised-feature-extraction scheme, a case study was conducted with data from the RK4 rotor kit. The results showed that the proposed scheme outperformed existing methods in terms of fault classification accuracy. We anticipate that the proposed scheme is promising as it can minimize the reliance of expert’s experience and domain knowledge.