Machine Fault Diagnosis (MFD) plays a vital role within the manufacturing industry. Faults can generally be detected by observing the changes that occur during their operation which causes the machine to deviate from its normal behaviour. The signs of trouble are manifested by changes in certain parameters such as temperature, sound and energy consumption. Faults from its inception will progressively get worse and if undetected will lead to catastrophic failure. By collecting operational data from an electrical machine under normal or healthy as well as faulty conditions an Artificial Intelligence (AI) based algorithm can be trained to detect faults. The focus of this research is on applying Deep Learning (DL) methods to the problem of diagnosing faults within electrical machinery. In this research Transfer Learning (TL) approach to detect rotor and bearing faults within an induction motor is proposed. A pretrained ResNet-34 Convolutional Neural Network (CNN) was constructed using the PyTorch Deep Learning library to automatically extract features from a number of input images such as infrared images, current signals, and vibration signals. The Gramian Angular Field (GAF) algorithm was used to encode current and vibration signals as images. A TL approach allows NN models to be trained significantly faster and with less data than a standard DL approach. The effectiveness of the different inputs was compared by building seven separate DL models and training each model on a different combination of the five different types of input data used to classify faults. It was found that the best performing model was the one that used only thermal images, achieving a mean test accuracy of 99%. This research also concluded that GAF algorithms are insufficient for capturing information on the health condition of an induction motor as they produced images that had low variance across fault conditions. Moreover, it was found that the ResNet-34 model was extremely susceptible to overfitting when trained on encoded current and vibration data due to relatively small datasets used in training. The main advantage of the proposed innovative use of TL approach is attributed to its faster training with less data.