Research Of Real-time Fault Detection Based On Neural Networks

In the process of fault detection and diagnosis, the relevant of fault information can be extracted from the residual signal, so the primary task of fault detection is to obtain the residual signal of the system, and the design goal of fault detection is residual. In the control system, Between the inputs、outputs and state variables of the system, there are some functional relationship, so we can use the variables measured to construct a residual generator, when the system has no fault, the residual of system is less than (assuming no external interference)the threshold, and when the system has fault, the residual will be greater than the threshold. This is the main idea of fault detection.In this paper, based on the basic idea of fault detection, use neural network as tool to design the residual generator, and with the Multi-function process control platform as the experimental platform to validate the proposed algorithm.And do some relerant research from the following aspects:1This paper present RBF-BP hybrid neural network model, according to BP neural network with self-learning, promotion and general ability, as well as RBF neural network with fuction approximation ability, the network model has the advantages of the BP network and RBF network。BP neural network can make up the limitation of RBF neural network, which the role function is not overall situation; and BP neural network with the lack of local optimum and the slow of convergence, the RBFneural network can make up it.For the training of the RBF-BP hybrid neural network model, the input of system as the input of neural network, and the output of system as the output of the network, at the same time, Using genetic algorithms to optimize the network, that is construct a simulation system instead of the real system. When the actuator of the system has fault, the output of simulation system is normal, and the actual output will be abnormal, a component of the residual will be greater than the threshold value, that is, εi>δi, at this time, believes the actuator of the system has fault.2For fault detection method based on state observers, the more traditional method is algebra in mathematics, such as LMI, but the algebraic method can result in repeat calculatation, and the neural network state observer can approximate any of the nonlinear function, so it study the aspect of BP network state observer in this paper.The traditional neural network state observer method estimates the state variables is according to the input and output of the system, who directly relate to it, but there are some functional relationship between the output and state variables for every sensor, so in this paper, for the estimation of state observer related to i sensor, the input of the network is the input of system and the output of the j sensor.3For control system with no external interference, use of neural network function approximation ability realize the simulation of real system, and generate the system residuals, and realize the fault detection and diagnosis by residuals, it can, but if the system has external interferences, the mathod can’t meet the requirements. Because if existence of unknown input, even if the sensor of system has no fault, the state error and residual signal also will deviate from zero, so wo must estimate the unknown input, and into the simulation system, stabilized. If there is no fault, the residual signal will be close to zero, otherwise, even if the state variable error will gradually approach zero, but the residual signal will deviate from zero too.