This work applies an unsupervised deep feature learning to finding patterns of interharmonics. The main objectives of this work are to provide an additional graphical tool to handle two distinct data inputs: (a) individual interharmonics components in time-series; (b) broadband spectrum by employing spectrograms. Both data inputs are analysed employing an autoencoder based on convolutional neural networks followed by clustering. The application of the method results in the most common patterns in time-series or spectrograms. Two study cases are presented by applying the method to measurements from solar installations in Finland and Sweden. The results show the usefulness of the method to recognize interharmonics in a single frequency and broadband spectrum.