The use of frequency tensors for representing discrete time series information through wavelet transformations provides a methodology that enables the application of machine learning methods in discrete spaces with displacements or states, and transitions. One example of this approach has been used in the detection of mobility patterns in a geographical area. Currently, this form of data processing is implemented as a continuous process in smart city platforms for mobility analysis applications. In this article, we propose normalizing the data structures used to conform to this type of frequency tensor. This technique was applied in a mobility cluster analysis conducted by the authors and will be useful for other machine learning applications. The integration of these data structures into machine learning algorithms was explored in an unsupervised learning setting. Our work presents an alternative method for constructing datasets that effectively transfer the necessary information to facilitate learning processes before the development of analysis models.