The existence of abnormal values in data sets of day-today actions is common. Usually denoted as outliers, these values are frequently excluded from the dataset. However, in fraud detection, disease diagnosis and other areas, outliers can often add relevant knowledge. In this paper, we present an application of a methodology to detect out-liers in multivariate time series by translating the multi-type variables into textual data (strings). After that, conventional classification algorithms can be applied to identify outliers. An application is performed into a real dataset extracted from the energy field. The proposed method was evaluated using a dataset that represents boiler operations and the main goal is to identify their faults. A major part of the work concerns the dataset processing steps that enable the application of common machine learning algorithms. It is worth mentioning that, besides the boiler malfunctions, normal operation cycles were also identified. We aim to improve the real-time fault identification of the operating devices allowing safer appliances.