In the multi-layer perceptron,its fault tolerance is highly related to its redundant hidden nodes. More redundant hidden modes will have the advantage of higher degree of fault tolerance but have the disadvantage of more computational and hardware complexities. In this paper we shall give an overall analysis of its hidden nodes and its fault tolerance. Falilt tolerance in the neural networks is first discussed. Then we propose a new strategy which can give the maxinum number of crashed nodes allowed within a given error for the neural network.In addition, we can remove some redundant hidden nodes for less computation. Simulation results are also given to illustrate the correctness of our method.