Electric railway systems (ERS) are characterized by several particularities regarding the return current circuits, moving loads, multiple sources of waveform distortion, and extensive deployment of static converters from various manufacturers, topologies, and solutions. This work presents a methodology for application of load monitoring to classify rolling stock (RS) waveform distortion signatures. The proposed methodology combines the benefits and advantages of unsupervised deep learning and reconstruction error performance classification for performing non-intrusive load monitoring (NILM) in ERS. It consists of adapting autoencoder-based novelty detection for load classification problems. The method is applied to pantograph measurements from four rolling stock items using two types of data input (harmonic spectra up to kHz and VI diagram images), which are compared in binary classifications of the same kind of railway electrification. The methodology shows suitable classification performance with high accuracy, scoring an average of 98.81 % for spectrum input and 97.77 % for VI diagram input. It has also been validated with a NILM dataset (LIT) for multi-class applications showing 99.13 % for spectrum input and 94.28 % for VI diagram input. The proposed method has suitable computational times and scalability, allowing application to a wide range of NILM and classification problems using distortion signatures.