The time-frequency map (TFM) is frequently used in condition monitoring, necessitating further processing to select an informative frequency band (IFB) or directly detect damage. However, selecting an IFB is challenging due to the complexity of spectral structures, non-Gaussian disturbances, and overlapping fault signatures in vibration signals. Additionally, dynamic operating conditions and low signal-to-noise ratio further complicate the identification of relevant features that indicate damage. To solve this problem, the present work proposes a novel method for informative band selection and local damage detection in rolling element bearings, utilizing non-negative tensor factorization (NTF)-based dependence map analysis. The recently introduced concept of the dependence map is leveraged, with a set of these maps being factorized to separate informative components from non-informative ones. Dependence maps provide valuable information on the auto-similarity of spectral content, while NTF, a powerful tool commonly used in image processing for feature extraction, enhances this process. The combination of these methods allows for the extraction of IFBs, forming the basis for local damage detection. The effectiveness of the proposed method has been validated using both synthetic and real vibration signals corrupted with non-Gaussian disturbances.