From past few years, accurate fault detection and localization are important for enabling efficient fault management in smart grids. Existing approaches for fault detection in power line communication systems for smart grids had face several challenges which include noise interference, complex network topologies and latency issues. Therefore, this research proposes One Class Support Vector Machine-based Autoencoder (OCSVM-based Autoencoder) for fault detection. Initially, the data is collected from smart grid power line communication dataset and this data is preprocessed by using Principal Component Analysis (PCA) which reduces the dimensionality. Then, features are extracted by using Empirical Mode Decomposition (EMD) with Hilbert transform and these extracted features are selected by using Gaussian mixture model and mutual information which handles complex distributions. After that, the detection of faults in smart grid in done by using OCSVM-based Autoencoder which effectively detected the faults in power line communication systems for smart grids. The proposed OCSVM-based Autoencoder achieved better results in terms of accuracy (99.01%), precision (99.00%) when compared with existing Random Forest.