The objective of this study was to develop a method for fault detection of an engine-out oxides of nitrogen (NOx) sensor. The aim of the developed method was not only to isolate a fault with the NOx sensor, but to also diagnose faults in other engine subsystems that may result in higher engine-out NOx production. The developed fault diagnostics are aimed at providing reliable, accurate determinations of sensor output, in-lieu of physical sensors. The data for the development of numerical models in this study was derived from in-use emissions data of a 2014 Freightliner equipped with a 2013 Cummins ISX15 engine. Data included engine control unit (ECU) data from a variety of vehicle operation in southern California that included interstate, highway, regional, local, and near dock locations. For this method of fault detection, a virtual sensor was created using an artificial neural network (ANN) with an input configuration using 12 engine parameters, which provided the most accurate results in this study. These parameters included engine speed, engine torque, fuel rate, intake temperature, boost pressure, exhaust temperature, coolant temperature, oil pressure, the first derivative of engine speed, the first derivative of engine torque, the second derivative of engine speed, and the second derivative of engine torque. The neural network could then be used to predict expected NOx values. The ANN NOx model was trained on a subset of the data and later validated with another subset of the available ECU data. Two different sets of training data, and seven validation data sets were used for prediction evaluation. The study also included the insertion of fault data and run against the model to test for fault detection with the best performing data set. It was found that the network is able to predict NOx within 1-5% at highway operation, when trained with highway data, enabling the detection of NOx sensor faults as well as faults in engine subsystems that were included in the input parameters for the neural network. Three different types of sensor failures, including a step, ramp, and square function failure, were implemented in the validation data, which caused an increase in error between the actual and predicted NOx production to increase between 15-200%, creating the means of detection.