Research Of Fault Diagnosis And Prediction For Boiler System Of Power Plant

The software of diagnosis and prediction for boiler fault is developed by using three-layer feed forward network and expert system technology. The L-M (Levenberg-Marquardt) method is added in this paper. The knowledge databases of fault diagnosis are added in the process of software development to fix the structure of network. Knowledge training module, simulation module, and interpretation module are programmed with Visual C++. Then the samples are trained with this network. During the training, the network parameters (such as circulation times, neuron number of hidden layer, step size) were modified to make sure that the training samples could be convergent to a relative small and stable value with a small learning error. At last, the simulation tests on drum excess water are carried out by using the fault diagnose software. Some comparisons between the BP networks and L-M networks can be made through sample training. And the result indicates that: the L-M method, which is the combination of Newton method and descent method, can avoid the problem of slow training speed and local minimum point in Bp networks when the weight number is small. And thereby increase network efficiency.The fault diagnose software has such a capability of simulating the speculation of human expert to a certain degree that it can diagnose the faults for boiler accurately. And the achievement of this paper can be used to power plant and imitate electrical machinery plant.