Neural network fault diagnosis system for a diesel-electric locomotive's closed loop excitation control system

In closed loop control systems fault isolation is extremely difficult due to the fact that if feedbacks are corrupted or actuators can’t produce a desired output, a system reacts due to an increase in error between the measured variable and the set input variable, which can cause oscillations. The goal of this project is to develop a fault detection and isolation system for the isolation of faults, which cause oscillatory conditions on a GE Diesel-Electric Locomotive’s excitation control system. The proposed system will illustrate the use of artificial neural networks as a replacement to classical analytical models. The artificial neural network model’s design will be based on model-based dedicated observer theory to isolate sensor, as well as component faults, where observer theory will be utilised to effectively select input-output data configurations for detection of sensor and component faults causing oscillations. Owing to the nature of the locomotive’s data acquisition abilities, the model-based observer design will utilise historical data to design an effective model of the system which will be used to perform offline sampled fault detection. This method is proposed as an alternative to trend checking, data mining, etc. Faults are thus detected through the use of an offline model-based dedicated observer residual generator. With the use of a neural network a number of parameters affect the accuracy of the network where the primary source of ensuring an accurate model is training. The project highlights and experiments with these parameters to ensure an accurate model is trained with the use of the gradient descent training algorithm. The parameters which are considered are learning rate, hidden layer neurons, momentum and data preparation. The project will also provide a literature review on residual evaluation techniques used in practice and describe and evaluate the proposed method to perform residual evaluation for this specific application. The proposed method for residual evaluation was based on two principles, namely the moving average, as well as the simple thresholding techniques. The developed FDI system’s performance was measured against known faults and produced 100% accuracy for the detection and isolation of sensor and components causing oscillatory conditions on the locomotive’s excitation system