| This dissertation describes model based approaches for Fault Detection, Isolation, and severity Assessment (FDIA) for a 20 HP, rotary, single-screw air compressor. A sensory subsystem was set up to measure physical variables which include the ambient temperature, radiator temperature, cooling oil entrance and exit temperatures, angular velocity, air mass flow rate, exhaust pressure, and motor current. In addition to these sensors, an optical non-contact tachometer was developed. Experimental values of sensor outputs from the compressor under various faulty conditions including radiator faults, gaterotor wears and increased frictions of different severity have been obtained.; With these measurements, the utility of linear modeling technique was established for fault detection and isolation. Linear mathematical models in the form of linear differential/difference equations, which describe the causal relationships linking physical variables, were derived from physical laws and were identified through system identification techniques. Using baseline models, a fault detection scheme that does not require prior experience of faults was developed to detect abnormalities. The frequency responses of the models were used to isolate the faults. It was shown that the use of model parameters, which frequently are influenced by faults more directly, has enabled early detection and accurate recognition of the compressor's internal faults.; However, it was also found that the operating range of the compressor exceeded the linearity of the model and, therefore, practical nonlinear modeling techniques were needed to detect, isolate and assess faults more accurately. Based on an existing nonparametric nonlinear model technique, a fault diagnosis and severity assessment method for the compressor was developed. The first order and the second order transfer functions of the compressor were identified from the measurements by the nonparametric nonlinear modeling technique. The transfer functions showed not only distinct patterns for the baseline and various faulty conditions, but also consistent trends that are well correlated with the extent of damage. The results clearly showed that nonlinear modeling offers a better accuracy and therefore more consistent diagnosis and severity assessment. However, the nonparametric approach requires a large set of measurement data and much computation power. A parametric approach is, hence, more desirable.; To facilitate development of a parametric nonlinear modeling technique, a fully automatic neural network structural and weight learning algorithm was developed. The Augmentation by Training with Residuals, ATR, requires neither guess of initial weight values nor the number of neurons in the hidden layer. The algorithm takes an incremental approach in which a hidden neuron is trained to model the mapping between the input and output of exemplars and it is then augmented to the existing network as training continues. The exemplars are then made to be orthogonal to the newly identified hidden neuron and are used for the training of next hidden neuron. |
