Adaptive Critic Networks for Real-time Fault Diagnosis in Electric Power Production

点击翻译按钮获取中文摘要

This project will develop a architecture for real fault diagnosis of electric power generating systems. It will focus on fossil fueled and nuclear power plants, characteristic of complex, large-scale systems amenable to analysis. The fault diagnosis architecture is a network of neural networks: At the signal processing level of the hierarchy, a recurrent multilayer perceptron network is used to recursively identify the nonlinear plant dynamics, whereas another neural network with similar structure is used for state and parameter estimation purposes. At the decision making level of the hierarchy, an adaptive critic network is used for detecting and isolating potential faults. The adaptive critic is composed of three neural networks, a faulty system model, a performance optimizer (critic), and a classifier (action) which selects and reports the time, location, magnitude, and nature of the faults. The resulting diagnosis procedure exhibits generic and therefore portable architecture. In the presence of real-world operating conditions characterized by noisy sensor and actuator measurements, as well as process noise and drifts, the proposed approach would be capable of detecting, isolating, and accommodating system faults which have not been encountered or anticipated during operation. It is anticipated that successful completion of the proposed project will significantly impact the electric power industry, and complex, large-scale nonlinear systems in general.