Power transformers are critical assets in the power system, and routine condition assessment of them is necessary to ensure reliable and cost-effective energy trans- mission. As a transformer operates, thermal and electrical stresses degrade its electrical insulation and remaining lifespan, and these degradation processes pro- duce gas compounds that are dissolved in its insulating oil. Dissolved gas analysis (DGA) assesses a transformer’s condition by measuring these gasses, and numerous interpretation techniques exist based on physical mechanisms and experimental data to determine the associated faults that caused the gassing. Artificial neural net- works (ANN) are a deep learning topology, which is a type of artificial intelligence (AI). ANNs are proven to be useful tools for determining complex relationships in data — particularly for classification purposes — and lend themselves suitable for DGA interpretation. Due to the dependence of AI models on their provided datasets, uncertainty regarding their accuracies when predicting new assessments is expected. This thesis explores the use of ANNs to interpret DGA for improved certainty in fault classification, and postulates that the existing DGA interpretation methods encode relationships that are not easily learned by ANNs. This work proposes a methodology that combines select methods (e.g., Duval’s Triangle 1, Duval’s Pentagon 1, and IEC Ratios) as input features along with a structure that also learns novel gassing-fault relationships in an attempt to improve performance and trust with DGA interpretation. Three models with varied input features are optimized with respect to their hidden-layer dimensions and training parameters, and are evaluated using repeated random subsampling with K-fold validation. The results from this work demonstrate that using ANN’s for DGA interpretation offers improvements over conventional methods — notably in low-temperature fault classification — and shows the best-performing model for the given dataset is achieved when industry methods along with gassing data are used as input features. Where the existing industry methods can conflict on assessments or provide unreliable results, the proposed models can offer a more compelling diagnosis of a transformer fault, and ultimately empower the asset’s owner to optimize its lifecycle.