Neural Network-Based Sensor Fault Diagnosis For A Nonlinear System

Fault detection and isolation for nonlinear systems, especially for systems with unknown nonlinear characteristics and unknown inputs, has been aroused more attentions in recent years. Various fault diagnosis methods have been developed. Studies show that neural networks are the most powerful tools to approximate almost every nonlinear system and have reliable accuracy. In this thesis, we investigate the fault diagnosis for a hydraulic system with severe nonlinearity under nine different operation conditions using BP neural networks. Sensors' fault are considered, which include abrupt parameter changes and slowly parameter drifting.First, the system operation conditions that include three input types (square, sinuous and triangle) with different frequencies (l/250hz, l/5hz and 1hz) are discussed. In fault detection processes, we construct 4 three-layer BP neural networks to learn the normal states of the system. The outputs of trained neural networks are as the estimate values of the sensors. Then the difference between the output of the trained neural networks and that of the actual system is used as a residual, which detects the sensor fault. The diagnosis can then be performed by further exploiting the information provided from the residual signals according to certain rules introduced in detail in the thesis.Simulating results show that the method can isolate any fault of the sensors in any input types and normal work frequencies. Unknown input is also considered. Any sensor's changes larger than their 3% normal values can be detected precisely.